Image Processing for Detection of Cataract, Retinopathy Of Prematurity and Glaucoma Arezoo Motamed Ektesabi Faculty of Science, Engineering and Technology Swinburne University of Technology A thesis submitted for the degree of Doctor of Philosophy 2015
254
Embed
Image processing for detection of Cataract, Retinopathy of ...€¦ · After image acquisition, the rst image processing stage is the image pre-processing. The general processes such
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Image Processing for Detection of
Cataract, Retinopathy Of
Prematurity and Glaucoma
Arezoo Motamed Ektesabi
Faculty of Science, Engineering and Technology
Swinburne University of Technology
A thesis submitted for the degree of
Doctor of Philosophy
2015
i
Declaration
This thesis is the result of my own work and to the best of my knowl-
edge, includes nothing, which is the outcome of work done in collabo-
ration except where specifically indicated in the text. It has not been
previously submitted, in part or whole, to any university of institution
for any degree, diploma, or other qualification.
Signature:
Arezoo Motamed Ektesabi
2015
I would like to dedicate this thesis to my loving parents,
Mehran Motamed Ektesabi & Sharareh Soufi Siavash
and my brother, Arman Motamed Ektesabi
In memories of Shirin Salimi Pirkouhi, my grandmother,
who was always inspiring me to continue my studies.
Acknowledgements
I would forever be grateful for all those who guided me and encouraged
me to challenge myself, never give up, advance and succeed in life.
A special thanks goes to my principal coordinating supervisor, Profes-
sor Ajay Kapoor, for his mentoring and support throughout my candi-
dature. It was with his continuous guidance, commentary, suggestions
and motivation that the completion if this thesis became possible.
I hereby would also like to acknowledge Associate Professor Richard
Manasseh, my coordinating supervisor, who at many times inspired me
and directed me to clarify my thought processes and aided me in my
decisions.
Throughout my candidature I received many invaluable supports from
many individuals and many friendships were formed. In particular I
would like to thank Dr Michelle Dunn who introduced me to image
processing.
I would like to thank my mother, Sharareh Soufi Siavash, who taught
me how to write and read prior to attending school; stood beside me
throughout my studies; and who showered me with love and encouraged
me to grow.
Many thanks to my father, Mehran Motamed Ektesabi, who introduced
me to the field of engineering from an early age; who was always there
throughout all the hurdles of life and was there when I needed an
advice; who believed in me, motivated and inspired me to progress and
achieve my best.
My parents, you are my first and long life teachers, my best friends
and mentors, I can never appreciate you enough for all that you have
done for me. Words cannot express how I feel about you. I just hope
you accept my sincere thanks and admiration.
I would also like to thank my younger brother, Arman Motamed Ek-
tesabi, who made me laugh when I was down and showed me his per-
spectives about the importance of life.
My grandparents, each in their own way, motivated me. I hope I have
done them proud, specially my grandmother, Shirin Salimi Pirkouhi,
who would have loved to see this day but unfortunately lost her battle
to cancer. Her dream was so that I could continue my studies and it is
with her well wishes that I have reached this far. May one day, I could
take part in research for early diagnosis of cancer.
My family and friends, my most valued treasures of life, I appreciate
each and every one of you for your positive encouragements and price-
less support. You have shown me how to live and taught me about
life’s vast opportunities. I am pleased to have had the opportunity to
know you and be part of your lives.
Publications
Book Chapter:
• A. Ektesabi, A. Kapoor, ”Fringe Noise Removal of Retinal Fun-
dus Images Using Trimming Regions”, Emerging Trends in Image
Processing, Computer Vision, and Pattern Recognition, Elsevier
Inc. Jan, 2015.
Conference Proceeding:
• A. Ektesabi, A. Kapoor, ”Exact Pupil and Iris Boundary Detec-
tion”, International Conference on Control, Instrumentation, and
Automation (ICCIA), Shiraz, vol. 2, pp. 1217-1221, 2011.
• A. Ektesabi, A. Kapoor, ”Complication Prevention of Posterior
Capsular Rupture using Image Processing Techniques”, Proceed-
ings of the World Congress on Engineering 2012 (WCE 2012),
vol. I, July 4 - 6, London, U.K., pp. 603-607, 2012.
• A. Ektesabi, A. Kapoor, ”Removal of Circular Edge Noise of Reti-
nal Fundus Images”, International Conference on Image Process-
ing, Computer Vision and Pattern Recognition (IPCV’14), Las
Vegas., 2014.
• A. Ektesabi, A. Kapoor, ”Optic Disk Localisation Using Con-
secutive Adaptive Thresholding Technique”, IEEE International
Conference on Image Processing (ICIP 2016), Arizona., 2016 -
Under Review.
Abstract
The field of ophthalmology is in need of more support as it is unable
to meet the need of the growing population. This thesis considers pro-
cedures which may be used as part of an assistive telemedical tool for
aiding ophthalmologists in diagnosing wide range of ophthalmological
disorders including Cataract, Retinopathy of Prematurity and Glau-
coma, which affect more than 60% of the population worldwide. Many
different image processing techniques have been analysed and in the
process some of the most favourable and advanced systems have been
selected for identifying some of the key features of the eye which are
commonly used by ophthalmologists for disease detection.
To address this aim and create a more suitable telemedical solution,
different stages of image processing is reconsidered and enhances in
the study. The stages include, image pre-processing, feature locali-
sation and feature extraction. The aim is to create simple, fast but
universal algorithms and procedures which could be implemented on
any captured data with any specifications.
After image acquisition, the first image processing stage is the image
pre-processing. The general processes such as the colour conversion
to the gray scale and green band selection, masking the region of in-
terest and preliminary filtering for sharpening the images are initially
implemented. However, to improve the results in further stages, new
procedures such as a trimming circle to reduce fringe noise and im-
age colour enhancements are also implemented. The final results show
significant improvements and more accurate findings in these cases.
The next stage is the feature localisation stage. Previous studies have
shown the main areas of interest in retinal images are vessels, Optic
Disk and the Macula. The features are extracted using the new pro-
posed algorithms. The results are promising and the localisation is
compatible with previously conducted studies. Moreover, in this stage,
another approach is suggested resulting in the Iris and Pupil localisa-
tion. The method may be used both for biometric purposes as well as
inter-operatively in surgeries such as those of cataract.
In the feature extraction stage, different methodologies are suggested
for detecting the centre of the Iris, Pupil, Optic Disk and Macula. The
radius and the area of these features are also calculated and compared.
For vessels an approach is suggested for detecting its end points. The
use of the information may result in detection of different diseases such
as Cataract, ROP and Glaucoma.
To further assist the ophthalmologists and medical practitioners an
approach is proposed which results in mapping of the retina, which
may then be used as an aiding tool for disease diagnosis, progression
and treatment.
Lastly, to reduce the error associated with each result, the light refrac-
tion within the eye is considered and the error calculated. The error
can then be taken under consideration while analysing each result.
The outcomes of the following study suggests a reliable yet cost-effective,
simple and fast approaches in which captured eye images may be anal-
ysed as part of an automatic assistive telemedical tool.
2.3.4.3 Manual Small Incision Cataract Surgery (MSICS)
Manual Small Incision Cataract Surgery (MSICS) is the most commonly used
technique in developing countries and it surfaced literature in the early 1990s [51].
In the MSIC process, the lens is removed as a whole through a self-sealing scleral
tunnel wound [52]. The wound does not require any sutures and is smaller than
the ECCE surgery, about 6.5mm [53].
Figure 2.9, represents the steps in Manual Small Incision Cataract surgery for
Cataract removal.
Figure 2.9: Manual Small Incision Cataract surgery [54]
MSICS is more cost effective, has faster rehabilitation and would result in bet-
ter visual acuity in comparison to the ECCE [21, 22, 35]. However, in comparison
to the Phacoemulsification technique the outcome of the surgically induced astig-
matism is higher in MSICS [21] and it may lead to several post-operative refractive
errors [22]. Overall the visual acuity of the phacoemulsification has proven to be
better in comparison to other available techniques [55, 56].
2.3.4.4 Phacoemulsification
Phacoemulsification is the most modern technique in Cataract surgery. It refers
to the procedure were the lens is divided into pieces and emulsified by an ultra-
sonic surgical handpiece. The pieces are aspirated out with the chamber fluid.
The anterior chamber pressure is kept constant via irrigation of the balanced salt
solution [57].
The ultrasonic surgical device currently used in phacoemulsification was first
introduced by Kelman in 1967 and has improved extensively since [58, 59, 60].
Figure 2.10, illustrates the Phacoemulsification surgery for Cataract removal.
27
2. MOST PREDOMINANT EYE DISEASES
Figure 2.10: Phacoemulsification surgery [61]
Some of the advantages of phacoemulsification technique [21, 55, 62] over other
available techniques could be as follow:
• Smaller incision size - 1.0 ± 0.12 mm
• Less invasive - Smaller incision size, no sutures
• Short surgical time - about 10 minutes
• Less surgically induced astigmatism
• Less leakage of fluids - Type and direction of the incision as well as the blades
used ensures that the anterior chamber fluid leakage is minimal.
• Rapid recovery
• Better visual acuity - Corrected vision as a result of lens replacement
The most recent advancements in phacoemulsification are that the needle
tip vibrates longitudinally and horizontally at frequencies ranging between 28-
50 kHz [59]. As a result the patients are exposed to low frequency ultrasonic
energy and the heat it may produce. It should also be noted that since the heat
28
2. MOST PREDOMINANT EYE DISEASES
may damage the corneal endothelium and hence affect the overall surgical out-
come, the exposure time and average power of the device should be monitored
constantly [60].
2.3.5 Monitoring Surgical Trainees
Majority of the surgeons’ today use computer based training tools as a preliminary
basis for performing Cataract surgeries. However, with only a few hours of training,
their level of expertise may not be sufficient when it comes to real life complications.
Moreover, majority of the performed surgeries are also subjective with no ref-
erence to a validated standard. They are based on the surgeons’ experience. The
limited experience or severity of the surgical complication could result in the life-
long side effects in patients’ life.
As a result it is essential to have a monitoring system to see the overall progress
of the surgeons. This system can also be used as an assistive tool to train and
guide the surgeons through the surgery. For the case of Cataract, the first step in
creating this device is to exactly locate the Iris and Pupil which has been further
investigated in the following chapters.
2.3.6 Importance of Iris and Pupil for Diagnosing Cataract
Based on the above factors, locating the Iris and Pupil during the surgery can
be used to determine whether a complication has occurred or not. This can be
achieved by studying the extent of the variation which occurs in shape of the Iris
and Pupil.
2.4 Retinopathy of Prematurity(ROP)
Approximately about 1% of the neonates are born prematurely, with a birth weight
below 1,500g, while roughly about 0.5% weight less than 1000g. The overall birth
rate is about 1 per 100 inhabitants per year [17].
Usually the premature infants’ retinas have underdeveloped vascularisation.
ROP is believed to affect the postnatal abnormal growth of these retinal blood
29
2. MOST PREDOMINANT EYE DISEASES
vessels, resulting in the formation of vascular shunts, retinal neovascularization,
and even tractional retinal detachment which in severe cases may lead to blind-
ness [17]. It is a disease which affects both eyes of the infants and in some cases
the effect may be irreversible and lead to blindness.
Figure 2.11 illustrates how the vasculature differs between the normal out-
growth and the patients with ROP [63].
Figure 2.11: Illustration of differences between normal and abnormal retinal blood-vessel development in the child with ROP.
ROP was first described by Terry in 1942-1943 as ”retrolental fibroplasia” [64,
65]. In the following 10 years, ROP was recognised as the largest cause of blindness
in developed countries and was growing in epidemic proportions.
Soon after, oxygen therapy 1 was recognised as the major cause of ROP and
hence the use of it was restricted [65]. As a result of this discovery, the incidence
of ROP decreased significantly. However, this adverse reaction was also associated
with an increase rate of morbidity and mortality in the premature infants [66, 67].
Therefore the oxygen therapy was once again brought in but supplemental oxygen
delivery to the premature infants was monitored carefully to main adequate blood
levels [68].
During 1980s and 1990s new treatment modalities such as vitamin E supple-
mentation, cryotherapy, laser photocoagulation and nursery light levels were stud-
ied and considered effective in reducing chances of occurrence of ROP [17, 65].
Even with the controlled oxygen level and the new treatments, the number of
1Oxygen therapy is the administration of oxygen for chronic or acute patient care.
30
2. MOST PREDOMINANT EYE DISEASES
infants with ROP has increased since then [69, 70]. This is mostly due to the
advancement in technology and hence the increased survival rate of very low birth
weight infants [70, 71].
2.4.1 Worldwide Effect of ROP
Throughout the years ROP occurrence has remained very high and one of the
areas of interest and research. This may be due to the fact that this disease affects
premature infants and in some cases leaves long lasting irreversible results. In such
cases, the patient may be severely visual impaired or even blind. These patients
will have to go through life with a condition which could have been easily avoided
or minimised if they were treated on time.
Despite the available treatments and research being conducted in the field of
ophthalmology, ROP still is known to be one of the major causes of blindness
in children in both developed and developing world [70, 72]. The proportion of
childhood blindness caused by ROP goes from 8% in high income countries to
40% in middle income countries. In Australia and New Zealand, every 1 in 10
premature infants develop severe ROP [73].
Retinal detachment is quite uncommon in children, accounting for only about
1.7% and 5.7% of all retinal detachments [74], but it is the cause of blindness in
ROP. In general, retinal changes which may be indication of regressed ROP, include
myopic changes, displacement of macula and retinal vessels, retinal folds, pigmen-
tary changes, incompletely vascularized peripheral retina, abnormal branching and
tortuous and telangiectatic vessels [75].
2.4.2 Risk Factors of ROP
There are many risk factors associated with ROP. With advances in the neonatal
care, the number of surviving premature infants has increased significantly, which
in some case may lead to development of ROP.
The low birth weight and low gestational age are known to be strong risk factors
of ROP, where the smallest infants are more likely to develop ROP.
Oxygen has been recognized as another significant risk factor since the 1950s.
31
2. MOST PREDOMINANT EYE DISEASES
However, the direct correlation of duration and concentration of oxygen with sever-
ity of ROP is not yet confirmed.
Other factors such as degree of illness, sepsis, blood transfusions, white race,
multiple births, and being born outside a hospital also increase the chances of
developing ROP [17, 65].
Socioeconomic factors and health care conditions of each country should be
considered while recognising the risks associated with ROP. Statistics have shown
that the occurrence of ROP is significantly increased in the developing countries
due to health care system and lifestyle in comparison to developed countries.
2.4.3 Classification of ROP
Once the patient is diagnosed to have ROP, to begin the treatment, the first step
is to classify the ROP. The studies have shown that more aggressive diseases are
located in the posterior section of the eye. Figure 2.12 represents the zones and
extent which are used to determine the classification of the ROP [17].
Figure 2.12: Classification of ROP for the left eyes [17]
The classification comprises of three parameters:
1. Location - zone of the disease in the retina [17]:
• Zone I is the posterior circle centred on the optic disc. Its radius is
about twice the distance from the disc to the centre of the macula. It
is defined as the most posterior location of disease.
• Zone II is a circle centred on the disc with a radius equal to the distance
to the nasal ora-serrate.
32
2. MOST PREDOMINANT EYE DISEASES
• Zone III comprises the remaining temporal crescent.
2. Extent by clock hours of developing vasculature involved
3. Severity - stage of the observed abnormal vascular response [76]:
• Stage I - mild abnormal blood vessel growth.
– No treatment is required and the child eventually may develop nor-
mal vision without further progression.
• Stage II - moderate abnormal blood vessel growth.
– No treatment is required and the child eventually may develop nor-
mal vision without further progression.
• Stage III - Severe abnormal blood vessel growth.
– Abnormal blood vessels formation towards the centre of the eye
instead of following the normal growth pattern along the surface of
the retina.
– Some infants may not need treatment and develop normal vision.
– Some infants who have certain degree of Stage III and ”plus dis-
ease1” need treatment to avoid retinal detachment 2.
• Stage IV - Partial detachment of retina.
– Treatment is required. The bleeding caused by scars of the abnor-
mal blood vessels cause traction and pulls the retina away from the
wall of the eye.
• Stage V - Complete detachment of retina.
– Treatment is required. If the eye is not treated, the child may have
severe visual impairment and even blinded.
1Plus disease is when the blood vessels of the retina have become enlarged and twisted. Thisindicates the worsening of the disease. Treatment may prevent retinal detachment. Prior to theformation of plus disease, significant vasoconstriction may be present.
2Retinal detachment occurs as a result of accumulation of the Sub-retinal fluid in the spacebetween the neurosensory retina and the underlying retinal pigment epithelium. It is classifiedinto Rhegmatogenous, Tractional and Exudative based on the mechanism of the sub-retinal fluidaccumulation [77].
33
2. MOST PREDOMINANT EYE DISEASES
2.4.4 Screening for ROP
The suggested examination time for the first visit is about 32-34 week of post
menstrual age and for the second visit is about 38-40 weeks post menstrual age [17].
Prior to screening the pupil of the eye needs to be dilated. To dilate the pupil
three times every 5-10 minutes eye drops are used.
At the time a nurse needs to be present to constrain the movement of the
infant and also look out for vital signs and clear airways, as Bradycardia due to
the oculocardiac reflex is a recognized to cause complication during the examina-
tion [17, 69].
During the screening process, follow-up and therapy the location, extent and
severity of disease are monitored and evaluated. The changes in the different
segment s of the eye, presence of persistent and dilated vessels in the retina are
monitored to see whether the treatment is needed [17].
Digital retinal wide-field imaging system is used to monitor and capture images
of the retina. Using the obtained data evaluation of the shape, degree of arbori-
sation, diameter of retinal vessels and estimate the severity of the disease even in
the absence of complete imaging has become feasible.
2.4.5 Treatment of ROP
Once the patient is diagnosed and is in need of treatment, photocoagulation ther-
apy or cryotherapy is recommended. Since the early 1990s, laser photocoagulation
has been used [78, 79, 80, 81] and is the preferred treatment method in comparison
to cryotherapy [17, 82, 83, 84].
Incidence of ROP have significantly reduced as a result of by better screening
and prophylactic cryotherapy or laser photocoagulation [75]. The treatments
have reduced the occurrence of blindness by approximately 25%; however, the
visual outcomes after treatment are often poor and patient may not have 20/20
vision.
The American Guidelines indicate the time to treatment has to be within 72
hours [17], but in some cases treatment should be provided without further delay.
These include patients with advance stage of the disease or those with zone I and
rapid progression disease.
34
2. MOST PREDOMINANT EYE DISEASES
Recently it has been advised to start early treatments to avoid rare potentially
blinding disease. Earlier treatment is now recommended for aggressive forms of
ROP, such as zone I and posterior zone II disease. In these cases the treatment
can occur as early as 30.6 weeks post menstrual age [17].
2.4.6 Importance of Retinal Vasculature for Diagnosing
ROP
As indicated several occasions in section 2.4, ROP is an ophthalmological disease
which is caused due to abnormal growth of the vasculature in the retina. The extent
of the damage of this complication is dependent on the screening and diagnosis time
as early detection reduces the possibility of severe complications and blindness.
Hence an automated monitoring system could be used as an assistive tool to
aid the technicians and medical practitioners in their diagnosis. This system can
be used in remote, rural areas as a preliminary diagnostic tool which distinguishes
the patients prone to ROP from the normal patients. Moreover, by further analysis
of the retinal vessels in cases where severe cases of ROP are detected, the system
may outline the regions of the retina which are affected and are in need of further
treatments.
In order to create this system, it is crucial to extract the exact location of
the retinal vasculature. This can be achieved by analyzing the fundus retinal
images using image processing techniques. In this study, new automated image
analysis approaches have been considered for vascular localization and key feature
extraction. More details are included in the consecutive chapters.
2.5 Glaucoma
In Greece in 400 BC, the term Glaucoma was first used by Hippocrates to describe
a dimming of vision. Many years later, in 1862, the pharmacology of Glaucoma
was first detected with the isolation of physostigmine from the calabar bean [85].
Glaucoma is now the second leading cause of irreversible visual loss and blind-
ness [1]. Due to asymptotic characteristics of this disease [86] and with the aging
population and health issues such as diabetes [87], the incidences of Glaucoma
35
2. MOST PREDOMINANT EYE DISEASES
remain high and an area of concern. Hence, to minimise vision loss in patients,
early detection and treatment of Glaucoma is essential.
It has been found that the cause of Glaucoma is the progressive loss of Retinal
Ganglion Cells and their axons.
This in turn causes morphological changes in the OD and visual field [35, 88,
89]. The initial signs observed are usually hemorrhage-associated retinal nerve fibre
layer defects. This is then followed by the visible changes of the OD, including the
thinning of the neuroretinal rim, pallor and progressive cupping of the OD. Often,
the visual field defects are detected at the later stages, where more than 40% of
axons are lost [35].
2.5.1 Worldwide Effect of Glaucoma
The leading cause of the irreversible blindness in the world is Glaucoma. It is also
the most common cause of blindness after Cataract. Worldwide, it has contributed
to the 14% of the blind population. Those accounts for about 70 million people,
from which 10% have been bilaterally 1 blinded by this disease [90].
2.5.2 Pathogenesis of Glaucoma
It is believed that Glaucoma damages the ganglion cell and its respective axons,
which comprise the Retinal Nerve Fiber Layer (RNFL) [90].
The progression of this damage results in asymmetric changes to the Optic
Cup (OC) and as a result visual field loss. Since there is no functional loss prior
to severe structural damage, up to 40% [35, 90], it is quite difficult to detect
Glaucoma early on in the disease progression.
The morphology of the defected RNFL appears to follow the normal structural
pattern of the retinal RNFL. The RNFL is usually striated. It radiates from the
OD and is thickest in the superior and inferior poles in comparison to the nasal
and the temporal poles. However, the Glaucomatous RNFL changes can present
as focal wedge-shaped defects of varying width radiating from the optic nerve head
or as diffuse loss of the striations in RNFL [90, 91]. Focal loss is often detected in
1Bilaterally means affecting both eyes
36
2. MOST PREDOMINANT EYE DISEASES
the superior and inferior nerve fibers as Glaucoma usually affects these regions.
2.5.3 Risk Factors of Glaucoma
In the literature, several factors have been found which may influence and increase
the possibility of occurrence of Glaucoma [92]. These factors include:
• Age - older people are more likely to develop Glaucoma
• High Intraocular Pressure (IOP) - leading cause of Glaucoma
• Ethnicity - African, Latino and Asian descendants are more likely to have
Glaucoma
• Family History of Glaucoma
• Diabetes - the chance of Glaucoma doubles in diabetic patients
• Myopia (shortsightedness) - changes the internal structure of the eye, in-
creadint the chance for formation of Glaucoma
• Extremely high or low blood pressure - deprives the eye from adequate blood
flow, affecting the the rate of oxygen and nutrients as well as the waste
removal from the eye
• Long term usage of Steroid/Cortisone - increases the IOP and so results in
Glaucoma
• Injury to the eye
2.5.4 Classification and Screening of Glaucoma
Diagnosis and early treatment of Glaucoma is essential in prevention of vision loss.
Prior to implementing the right treatment method, the exact type of Glaucoma
has to be categorised.
There are several different types of Glaucoma. Some of which includes [90]:
1. Primary Open Angle Glaucoma - gradual increase in IOP
37
2. MOST PREDOMINANT EYE DISEASES
2. Normal Tension Glaucoma - known as Low Tension Glaucoma and occurs
when there is a progressive damage to the optic nerve under normal IOP
3. Angle Closure Glaucoma - inherited
4. Acute Glaucoma - sudden increase in IOP
5. Pigmentary Glaucoma - type of an inherited Open Angle Glaucoma
6. Trauma related Glaucoma - acute or chronic development as a result of an
injury to the eye
7. Childhood Glaucoma - occurs in children when there is an abnormal increase
in the IOP
2.5.5 Treatment of Glaucoma
As mentioned earlier, Glaucoma may cause an irreversible blindness, therefore
early diagnosis and treatment of it could be crucial to manage this disease. De-
pending on the severity of the Glaucoma, several different treatment options are
available [90], including:
• Eye drops
• Medication
• Surgery - Traditional or Laser
2.5.6 Importance of Optic Disk and Macula for Diagnosing
Glaucoma
For years, clinical approaches were used for monitoring patients with Glaucoma.
The ophthalmologists considered OD and its variation in shape to monitor the
progression of this disease. However, due to limitations of the subjective nature
of the evaluation and progression of the disease, the use of computerised image
analysis technique is suggested.
38
2. MOST PREDOMINANT EYE DISEASES
The automated assistive tool can aid in localising and extracting the infor-
mation from the OD and Macula. The obtained information can then assist in
diagnosis and prognosis of diseases such as Glaucoma. Further details in regards
to the image processing procedures involved in localisation of the OD and Macula
is covered in the consecutive chapters.
2.6 Summary
This chapter has covered the importance of vision in humans life. The field of
ophthalmology and some of the most common ophthalmological complications were
also discussed.
Three of the major leading causes of impairment in vision were investigated
in details, including the Cataract, ROP and the Glaucoma. The key features in
diagnosing these diseases have been defined and will be examined in more details
in the coming chapters.
The key features that have been found for Cataract, ROP and Glaucoma are
Iris and Pupil, Retinal Vessels, and OD and Macula respectively.
39
CHAPTER 3
IMAGE PROCESSING IN OPHTHALMOLOGY
Introduction Literature Review
Most Predominant Eye Diseases Image Processing in Ophthalmology
Thesis Outline Conclusion
Figure 3.1: Chapter Three Outline
3.1 Ophthalmological Complications
For years, diagnosing of ophthalmological disorders was being performed by obser-
vation. The results were very subjective and could have varied based on individuals
perspective and experience level.
In recent years, with advancements in biomedical applications and in particu-
lar image processing, new procedures have been implemented to provide a more
objective review of diseases and their diagnosis. To gain a better understanding
the current procedures, this reviews the available technology and image processing
methodologies.
Since the field of ophthalmology is quire broad and covers a wide range of infor-
mation, an in depth review of some of the main ophthalmological complications has
40
3. IMAGE PROCESSING IN OPHTHALMOLOGY
already been conducted and covered in the previous chapter. The diseases include
the Cataract, ROP and Glaucoma. The key features of interest in recognising
these diseases were also identified, including the Iris and Pupil, retinal vessels, OD
and Macula.
3.1.1 Importance of Image Processing in Ophthalmology
For years, health care system was only progressing based on the experiences and
knowledge of the health care professionals. However, with an increasing popula-
tion, longer life span and technological advancements, there is a need to change
the traditional methods of manual patient examination with more modern semi-
automated or automated procedures. This could be beneficial for both the patient
and the medical experts especially in regions where the number of experts are
much less than the number of patients.
Incorporation of the medical field with engineering, has led to a new field of
biomedical engineering. Biomedical engineering has played significant role in all
stages of medical procedure, including the prognosis, detection, treatment and post
treatment. This collaboration has led to increasing number of successful cases.
One of the main areas which has helped majority if not all the medical fields
significantly is imaging. With the advancements in imaging devices, nowadays
many of diseases and complications may be detected early on, leading to less
severe cases and early treatments.
Despite these significant life changing outputs, there is still much more to be
done and imaging continues to be a growing field.
Ophthalmology is also benefited significantly from imaging devices. Similar to
other medical fields, imaging has helped ophthalmologists in their prognosis of dis-
eases and their progression, detection of complications, inter-operative procedures,
post treatments and many more. It has also allowed researchers to have better
knowledge and view of the underlying structures of the eye, and its complications.
Imaging consists of different sections. The foremost step is the image capturing.
It is important to consider the requirements and the purpose of the image; based on
these specifications the image can then be obtained. Once the image is captured,
the image may be analyzed manually by the ophthalmologists. However, in many
41
3. IMAGE PROCESSING IN OPHTHALMOLOGY
cases the analysis by an expert may not be possible.
With the increase in world population and limitations in the experts and ad-
vanced resources, the manual analysis of the captured images may not be possible;
as a result a new field of telemedicine is introduced. Telemedicine is especially in
use in growing and developing countries and remote locations of developed coun-
tries such as Australia. This is where the images and preliminary analysis is done
remotely and automatically.
This new field is very much dependent on the collaboration and close work of
medical professionals and biomedical engineers. The knowledge of the engineers in
image analysis and the experiences of the medics have allowed the image processing
to be achievable and of great importance even in the field of ophthalmology.
The flowchart 3.2 illustrates the steps undertaken in image processing.
Figure 4.4: The difference between the view angle of normal angle, narrow angleand wide angle fundus cameras.
69
4. IMAGE ACQUISITION AND FUNDUS MAPPING
Fundus mapping is a more time consuming approach in comparison to the use
of wide angle of view cameras. However, the main advantage of this technique is
that the wide angle view is obtained using more magnified images, revealing more
details to the ophthalmologists. This in turn may improve the accuracy of disease
diagnosis and treatment despite.
Moreover, the availability and cost effectiveness of the normal angle of view
cameras make the fundus mapping more feasible and desirable especially in devel-
oping regions.
Since in both developing an developed countries the typically used cameras are
the normal angle of view, it is advised to use images from normal angle of view
cameras for further investigation. As it can be seen in Figure 4.5, using multiple
images and combining the results expands the field of view and be more useful
than just a single image when it comes to disease diagnosis.
Figure 4.5: Importance of fundus mapping
In this section, merging multiple images in order to obtain a wider view of the
retina from the typically used 30◦ fundus camera has been considered.
This would be quite different to those of the previously performed montage
models represented in the literature. In general, montage model is thought to be
a time consuming procedure and difficult to perfect. There seems to be problems
with presence of artefacts to the montage images.
70
4. IMAGE ACQUISITION AND FUNDUS MAPPING
To address these problems the following procedure has been suggested.
4.3.1 New Proposed Technique for Fundus Mapping
The objective of this section is to merge multiple fundus images together in order
to attain a wider FOV of the back of the eye. To do so, the geometric properties
or the eye is studied and the following method is introduced.
It is known that the back of the eye is curved; this property may be used to
introduce the following equations representing its horizontal curvature character-
istics. Moreover, in Figure 4.6 a geometric approximation of the retinal image is
illustrated.
Figure 4.6: Geometric representation of the proposed method for merging multipleretinal images. Radius of the Curve (R), Central Angle of the Curve (∆), CordLength (C), Tangent Length (T ), Middle Coordinate (M), External Distance (E)and the Middle (PM), Left (PL) and Right (PR) points can be viewed in theimage.
The Tangent Length, T , may be represented by the Equation 4.1. In this
equation, the Radius of the Curve is represented by R and the Central Angle of
the Curve in degrees (◦) is represented by ∆.
T = R× tan(
∆
2
)(4.1)
71
4. IMAGE ACQUISITION AND FUNDUS MAPPING
The External Distance, E, can be shown as:
E = R
(1
cos(
∆2
) − 1
)(4.2)
Moreover, the Cord Length, C, is:
C = 2Rsin
(∆
2
)(4.3)
The Middle Coordinate, M , can be defined as:
M = R
(1− cos
(∆
2
))(4.4)
Lastly, L, which is the Curve Length Distance between PM to the V ertex
(Right angle triangle to T and R) can be written as:
L =R∆π
180(4.5)
Since in capturing the fundus image, the device and its specifications are known,
the Central Angle of the Curve in degrees (∆) would also be known. For example,
for the Fundus camera with 30◦ FOV, the Central Angle of the Curve for each
image would be 60◦ based on the inscribed angle theorem.
As a result of this, the Cord Length, C, would also be constant for all the
captured images. This agrees with the observation that all the retinal images
obtained from the same device with same setting appear to have the same shape
and diameter. Therefore, calculating the diameter of the fundus images, would
define the Cord Length value.
Based on these, the Radius, R, may now be calculated by re-arranging the
Equation 4.3 and resulting in Equation 4.6.
C = 2Rsin
(∆
2
)R =
C
2sin(
∆2
) (4.6)
72
4. IMAGE ACQUISITION AND FUNDUS MAPPING
The equation found for the Radius, R, may now be substituted back into the
previously defined equation. Substituting Equation 4.6 into Equation 4.1 would
result in:
T = Rtan
(∆
2
)=C
2
(sec
(∆
2
))(4.7)
E, the External Distance, may now be defined by substituting Equation 4.6
back into the Equation 4.2:
E = R
(1
cos(
∆2
) − 1
)
=
((C
2sin(
∆2
)cos(
∆2
))−( C
2sin(
∆2
))) (4.8)
To simplify this further, the double angle formula may be used:
2sin(θ)cos(θ) = sin(2θ) −→ C
2sin(
∆2
)cos(
∆2
) =C
sin(∆)
Continuing on Equation 4.8, the External Distance, E, may now be:
E =
((C
sin(∆)
)−
(C
2sin(
∆2
)))
= C
(csc (∆)− 0.5csc
(∆
2
))(4.9)
Obtained result in Equation 4.6 can also be substituted into the Equation 4.4,
resulting in:
73
4. IMAGE ACQUISITION AND FUNDUS MAPPING
M = R
(1− cos
(∆
2
))=C
2
((csc
(∆
2
))−(cot
(∆
2
)))(4.10)
Similarly, substituting the obtained result in Equation 4.6 into the Equa-
tion 4.5, would now result in:
L =R∆π
180
=C∆π
360sin(
∆2
) (4.11)
From this it can be said that, all the required properties for approximating the
retinal horizontal curvature can now be calculated.
It is now time to merge multiple of these images, increasing the FOV. To merge
multiple images, it is best to have some overlapping regions. The overlapping
regions ensure that the possible artefacts which may have been formed due to
inaccurate positioning in the result are reduced.
Figure 4.7: Approximation of retinal curvature using the Middle Coordinate
74
4. IMAGE ACQUISITION AND FUNDUS MAPPING
The number of markers can also be increased as the OD and macula can be
detected. Using these markers in conjunction with the normal vasculature mark-
ers, the fundus map can be created with higher accuracy. Moreover, using the
approximation of the retinal curvature found via considering the Radius and the
Middle Coordinate of the eye, a 3D effect can be given to the outcome. Figure 4.7
illustrates this concept.
4.3.2 Implementation and Discussion
In order to map the fundus, it is important to define markers on each image so
that they could be used to overlap the image.
The advantage of the proposed approach over the readily available method-
ologies is that in the suggested case, the number of the markers has increased by
inclusion of the location of other retinal features. In the previous techniques, the
markers have been usually set based on the localized vessels. However in this case,
in conjunction with the localized vessels, other key features of the retinal image
including the location of OD and macula have also been used as markers.
Increase in the number of markers ensures that the error in creating the fundus
map is decreased and the overall image does not contain any duplicate images,
reducing the unwanted artefacts.
Moreover, the overlapping region has to be present, so that the combination of
the images could be achieved. Without the overlapping region, the images cannot
be placed next to one another as their location and the direction may be unknown.
The greater the overlapping region improved the accuracy of the fundus map,
but it also increases the computational complexity, reduces the speed of mapping
and also increases the need for using more images to create the full view of the
fundus map.
Based on this, it can be said that if time permits, fundus mapping could be
applied and used in developing or regional areas where the available resources
are limited. The advantage of this approach is that with minimal information;
knowing the fundus camera angle and the cord length of the taken images; a
simple yet reliable process can be applied. Moreover, since the number of markers
has increased, naming the location of the retinal vessels, OD and Macula, the
75
4. IMAGE ACQUISITION AND FUNDUS MAPPING
accuracy of the fundus mapping has also increased.
4.4 Refraction Studies
In image acquisition and interpretation, the light characteristics and effects are
of great importance. In ophthalmological instrumentations, many different light
sources, settings and angles have been considered for capturing images. Based on
the results and the desired outcome, the best light characteristics has then been
chosen and applied to obtain images.
However, once the images have been captured, the effects of the light beams
on the accuracy of the results in the interpretation stage have not been considered
in many studies. It is essential to know more about the light characteristics when
analyzing the results as it directly affects the accuracy of the calculations as it is
one of the main variables. In this section, the light refraction and how it effects
the overall interpretation of results is reviewed and studied.
The light beams tend to refract when leaving a matter and entering another
matter with a different density values, which are commonly known as refractive
index values. Therefore, when studying the light, considering the light refraction
based on these refractive index values are crucial and many studies have missed
this in their interpretations.
Air
Eye
θ2
θ1
Equation 4.12 indicates the relation between the index value of the angle of
incidence and refraction when light passes through two different materials.
n1sin(θ1) = n2sin(θ2) (4.12)
76
4. IMAGE ACQUISITION AND FUNDUS MAPPING
Where n1 and n2 refer to the refractive index values of the materials where the
light leaves and enters respectively. θ1 is the angle of incidence of light and θ2 is
the angle which the refractive ray created with the normal.
When light enters a material with higher refractive index, the angle of refraction
will be smaller than those of the angle of incidence, and hence the light will be
refracted towards the normal of the surface. However, if the refractive index of the
material is smaller, the refractive angle will be larger and light will be refracted
away from the normal.
n1 > n2 −→ θ1 < θ2
n1 < n2 −→ θ1 > θ2
4.4.1 Light Refraction In Retina
Based on the studies conducted by Hecht et.al. [143], the internal components of
the eye each have their own refractive index, hence the angle in which the light
enters the eye will not be the same as those reaching the back of the eye. As a result,
studying and implementing refractive index should be taken into consideration
while capturing or studying images.
Figure 4.8: Average light refraction indices for different regions of an eye.
77
4. IMAGE ACQUISITION AND FUNDUS MAPPING
Using the Equation 4.12, the angle of refraction of the light as it passes through
each sections of the eye can then be calculated. The results can be viewed in
Table 4.1.
Table 4.1: Refractive Index of the light passing through different regions of theeye.
N1 1
N2 1.376
N3 1.336
N4 1.406
N5 1.337
N1N2
0.726744186N2N3
1.02994012N3N4
0.950213371N4N5
1.051608078
All 0.747943156
Using the results from Table 4.1 in the Appendix, the comparison between the
calculated angle of the refraction and the expected incident ray over 180◦ and 90◦
has been plotted and can be viewed in Figures 4.9 and 4.10 respectively.
Figure 4.9: Comparison of incident ray and refractive ray - 180 degrees
78
4. IMAGE ACQUISITION AND FUNDUS MAPPING
Figure 4.10: Comparison of incident ray and refractive ray - 90 degrees
Figure 4.11 is an example of how the incident ray enters the eye and the refrac-
tive ray reaches the back of the retina. As a result of the difference in refractive
indices for each region of the eye, the bending of the ray is visible. The differ-
ence between the actual location of the ray and the expected location can also be
viewed.
Figure 4.11: Example of bending of the refractive ray in the eye
79
4. IMAGE ACQUISITION AND FUNDUS MAPPING
The graphs clearly show that there is a significant difference between the inci-
dent ray angle and the refractive ray angle. This may suggest that the expected
location of ray will differ from the actual location of the ray reaching the back of
the eye.
In order to statistically determine the significance of the results obtained for
the refractive ray in comparison to those from the incident ray, in this section the
Analysis of Variance (ANOVA) has been performed and shown in Table 4.2.
Table 4.2: ANOVA of the incident and refractive rays for 0-90◦ range
Summary
Groups Count Sum Average Variance
Incident Ray 19 855 45 791.6666667
Refractive Ray 19 554.3631598 29.17700841 261.1563218
ANOVA
Source of Variation SS df MS F-value P-value F-crit
Between Groups 2378.497 1 2378.487 4.518 0.040 4.113
Within Groups 18950.814 36 526.411
Total 21329.301 37
In the Table 4.2, for each group of results, incident ray and refractive ray
over the range of 0-90 degrees, the number of variables (count), their overall sum,
average and variance have been calculated and displayed in the summary section.
Moreover, for comparing the results using ANOVA, the Sum of Squares (SS),
Degrees of Freedom (df), Mean Squares (MS), the calculated F-value, P-value and
critical F-value (F-crit) have also been calculated and presented in the ANOVA
section.
The p-values are commonly used to determine whether the null hypothesis
could be accepted and rejected. The null hypothesis in this case is that there is
no significant difference between the incident ray and the refractive ray and the
study is to prove whether that is true or not. Depending on the p-value this could
80
4. IMAGE ACQUISITION AND FUNDUS MAPPING
be achieved.
P-values could mean:
• p ≤ 0.01 :very strong presumption against null hypothesis
• 0.01 < p ≤ 0.05 :strong presumption against null hypothesis
• 0.05 < p ≤ 0.1 :low presumption against null hypothesis
• p > 0.1 :no presumption against the null hypothesis
Based on the results illustrated in Table 4.2, the p-value is 0.040 which is less
than 0.05 but greater than 0.01. This means that there is a strong presumption
against the null hypothesis of no statistical significance between the two data sets.
This means that the two sets are significantly different and so when analytically
studying the images, the refraction of the light and its effect should also be con-
sidered.
Furthermore, the Fisher’s test (F-test) has also been found. The statistical
F-test determines whether the F-distribution is true under the null hypothesis.
The following is the formula for the one-way ANOVA F-test statistic:
F − test =Explained Variance
Unexplained Variance(4.13)
Using the Equation 4.13, the F-value has been obtained and as shown in Fig-
ure 4.2, it can be seen that the F-value is 4.518, which is slightly greater than
the critical F-value of 4.113. This suggests that the results may be significant at
the 5% significance level. Therefore, the null hypothesis can be rejected, suggest-
ing that there is strong evidence that the expected values in the incident ray and
refractive ray differ. This agrees with the results found for P-value.
Based on the above observations and results, it can be concluded that there
is significant different between the angle of incident and those of the refractive
angles and as a result should be taken under consideration when analysing the
outcomes. This may be beneficial to surgeons in their diagnosis of diseases as the
approximate location of the retinal features could be more accurately calculated
and determined using the angle of incident.
81
4. IMAGE ACQUISITION AND FUNDUS MAPPING
4.5 Summary
In this chapter the preliminary step of image processing was considered. The
image acquisition, light refraction and improvements to the field of view using
fundus mapping were the main areas of interest in this section.
Different factors for device and light source selection were considered in order
to highlight their importance in image acquisition and its impact to the overall
outcomes of the project. After due consideration, the eye and fundus images
used in this case study were obtained from online sources, captured from wide
range of devices with different settings. The main sources were the DRIVE and
STARE databases. The images were chosen to test the flexibility of the suggested
methodologies and determine the accuracy of the obtained results.
Furthermore, the impact of fundus mapping and light refraction has been in-
vestigated. In the image capturing, the effect of light refraction is significant and
therefore has been carefully studied in this chapter. The results have shown that
there is a significant difference in the incident and refractive rays and therefore
the variation has to be considered in order to aid the medical practitioners by
detecting the actual location of the key features of the retina.
Diagnosis and treatments of the retinal diseases can also benefit from wider
view of the retina, using fundus mapping. The use of multiple images from nor-
mal 30◦ angle of view retinal fundus images have been considered to create the
retinal fundus map with the proposed approach. The accuracy mapping has been
increased by using multiple different markers, including the location of the vessels,
OD and macula.
82
CHAPTER 5
IMAGE PRE-PROCESSING
Introduction Literature Review Thesis Outline Conclusion
Image Acquisition Image Pre-Processing
Implementation:
— Colour Separation
— Masking ROI
— Filtering/Noise Removal
— Image Sharpening
Further Modification:
— Contrast Enhancement
— Trimming
Feature Localisation Feature Extraction
Figure 5.1: Chapter Five Outline of Image Processing Stages
5.1 Overview
Improvements in the field of ophthalmology are indebted to advancements in im-
age capturing procedures and instrumentation. Previously, the visual inspection
83
5. IMAGE PRE-PROCESSING
of the eye was the only source for disease detection and treatment. However, en-
hancements in imaging and its processing significantly changed these traditional
approaches.
Implementing image processing procedures would require image and data. Ma-
jority of the times, the raw data would not be sufficient for prognosis of the dis-
ease. As a result further processing would be required. Image processing can
provide more information via analysing the outcome and detecting information
which might be missed by visual inspection. In order to do so, the image has to
be prepared and modified in the pre-processing stage.
In the pre-processing stage the acquired images are manipulated and noises
removed in order to enhance the speed of detection and results obtained in the
consecutive stages of feature detection and extraction.
Despite the rapid technological progression and knowledgebase understanding
of the eye structures and the underlying processes; in many regions especially
the developing countries, current resources may still not be available. On many
other occasions, the capturing devices may not produce high quality images or the
obtained images, maybe too noisy. Hence, it is important to filter images while
preserving critical information.
Consequently, the accuracy of the images and their readability may be affected,
further resulting in poor study of the patients’ health and imprecise disease de-
tection. It is of great importance to ensure that the readily available resources
and obtained results are well prepared for further prognosis by experts in the best
possible timely manner.
As a result, it can be said that the preliminary stage of image pre-processing
and modification plays an important role in disease detection. In this chapter, the
pre-processing stage has been considered.
The procedures include colour separation of the captured images, masking
the ROI, filtering and noise removal of the images and sharpening them. The
performed procedures ensure that the ROI is accurately detected and the overall
precision of detection is enhanced.
Furthermore, new modifications including the contrast enhancement and trim-
ming regions for ROI is also suggested. The trimming regions are defined so that
the errors associated with the localisation of the ROI, such as the OD is reduced.
84
5. IMAGE PRE-PROCESSING
Moreover, for betterment of the feature localisation, contrast enhancements are
suggested to be used.
5.2 Image Manipulation
Fundus retinal images captured are usually coloured images. Based on the work
conducted by Gonzales et.al. the coloured images are best to be converted into
either indexed or RGB (Red, Green and Blue) images [144, 145].
Gray scaled images have proven to reduce the complications and processing
time significantly. Therefore, the first step in image manipulation would be gray
scaling the RGB image. An example of such transition can be viewed in Figure 5.2,
part (a).
(a) Gray Scale (b) Red (c) Green (d) Blue
Figure 5.2: Colour band separation of a coloured image with respected histograms
The coloured image can also be separated in to its primary components of
85
5. IMAGE PRE-PROCESSING
the red, green and blue. The obtained results and their respected histograms are
displayed in Figure 5.2.
Figure 5.3: Colour component separation of RGB image in horizontal direction
Observing the results in Figure 5.2 and those in Figure 5.3, suggests that the
red channel of the RGB image is saturated, while the blue channel is empty.
Therefore, for the purpose of this study, the green channel is chosen for further
investigation.
This result agrees with the previous findings in the literature [115, 117, 146].
Similarly, Al-Rawi et.al. [147] conducted a study to determine the performance of
each of the colour bands in the DRIVE database by plotting the Receiver Oper-
ation Curve (roc) on an improved matched filter. The results indicated that the
average roc area for the red, green and blue bands were 0.9348, 0.9352 and 0.9339
respectively, once again suggesting that the green band is the most appropriate
channel for digital retinal imaging.
Based on this finding, all the coloured images in this study have been grey
scaled and their green channel have been chosen and used for further processing.
Some sample results are included as part of Appendix B.
5.3 Masking
After deciding on the channel of interest, which is the green channel of the image,
the region of interest (ROI) needs to be defined using a mask.
86
5. IMAGE PRE-PROCESSING
In retinal images, the main area of interest is surrounded by a black region.
Figure 5.4 illustrates a possible mask which could be used to define the ROI for
the sampled image. The white region in the mask is the ROI, while the black
region is the regions which are not of interest in the study.
(a) Image (b) Mask
Figure 5.4: Example of a possible mask for the sampled image
In the literature, there are many examples of how to mask the ROI. In many
studies, this mask is manually or automatically pre-defined and used for all images.
In other studies, methods such as the Otsu Method [148, 149] or Circular Hough
Transform [150] have been used to define the mask.
5.3.1 Otsu Method
Otsu method is based on the discriminate analysis and was first proposed by
Otsu in 1979 and since then was widely used in image processing applications.
Otsu method finds the optimal threshold in an image by thorough search of pixel
intensities for maximising the between class variances [148].
In the Otsu method, the image is separated into two classes of ”Object” and
”Background”, represented as C0 and C1 at the grey-level t.
Respectively, the within class variance 1, between class variance 2 and the total
variance are σ2W , σ2
B, σ2T . Based on the Otsu method, in order to find the optimum
1Within class variance is the weighted sum of the variances for each cluster.2Between class variance is the difference between the total and the within class variance.
87
5. IMAGE PRE-PROCESSING
threshold, one of the following functions with respect to t should be minimised
[149].
λ =σ2B
σ2W
η =σ2B
σ2T
κ =σ2T
σ2W
Since η is the simpllest of the three equations it is usually chosen and so the
optimal threshold is defined as:
t = ArgMin(η)
η =σ2B
σ2T
κ =σ2T
σ2W
Otsu method has been implemented and the results can be viewed in part (a)
of Figure 5.1.
5.3.2 New Technique for Masking Using Thresholding
In the study, a similar approach to the Otsu method is suggested and implemented.
Since fundus images are obtained using different devices with different settings, a
universal adaptive approach is needed, where the ROI could be defined for each
individual image, regardless of the capturing device settings. Since each device
setting is unique, for a universal automated process, the images obtained have
to be individually analysed and therefore each image would need to be masked
separately in order for its ROI to be defined.
The suggested method is an adaptive thresholding technique. It is quite fast,
reliable and easy to perform. The first step is to obtain the intensity of the image
and plot the histogram of the plane of interest, which in this case is the green
plane.
Studying the histogram closely reveals that there is a large peak at the lower
pixel intensities, which suggests presence of a significant dark region in the image.
Since the surrounding region is coloured black and the ROI is lighter, defining that
88
5. IMAGE PRE-PROCESSING
region and masking it would result in separation of the two regions.
To define the mask for the background region, a threshold is set where the
first major minimum occurs in the smoothed histogram. An example is shown in
Figure 5.5.
Figure 5.5: Histogram used to determine a threshold for masking the ROI
Once the threshold is set, all the pixel values in the image which have the pixel
intensities below the defined ROI is set to ”0”, and any values above it, is set to
”1”. The result is the creation of a binary mask, defining the ROI. The mask is
smoothed out by removing or filling up any noise which might appear as black
”holes” in the image.
An example of the proposed mask is shown in Table 5.1. In the figure, the re-
sults obtained using the proposed technique is compared with the results obtained
by implementing the Otsu method.
89
5. IMAGE PRE-PROCESSING
Table 5.1: Comparison of the masks formed by Otsu method and the suggestednew Thresholding method.
Technique Mask Masked Image
Otsu Method
New Thresholding Method
The comparison suggests that the proposed method is as accurate as the pre-
viously suggested Otsu technique. This approach has been implemented on more
than twenty images and the results indicate an accurate localisation of ROI for all
cases. The obtained masks have been included in the Appendix C.
Furthermore, in cases such as this one, where the two clusters are easily distin-
guishable, simpler yet reliable approach of thresholding is desirable. The suggested
approach is also faster and computationally less complicated in comparison to the
Otsu method as it only considers the occurrence of first major minimum instead
of calculating the minimised variances of different sections of the image. There-
fore the proposed technique can be used to define a mask for ROI as a universal
automated approach.
5.4 Filtering
Despite the presence of hardware filters, the obtained images are not ideal and
are still noisy. Therefore, it is essential to filter images and minimise noise prior
to any further processing. Since the used images are from open source databases
and so no further hardware filter may be implemented. Moreover, software filters
90
5. IMAGE PRE-PROCESSING
can be implemented on all images irrespective of the specifications of the image
capturing devices. Therefore, the study concentrated on software analysis of the
images and this section software filters have been applied and the details are as
follow.
5.4.1 2D Fast Fourier Transform
To enhance the processing time and reduction in computational complexity, the
2D Fast Fourier Transform (FFT) is suggested to be used as a filter. FFT is
computationally simpler because the filter is multiplied in frequency domain, while
in spatial domain it would have to be convoluted, therefore the FFT would result
in faster response time. It has been implemented and used as the basis of multiple
upcoming stages in this thesis.
FFT is an important tool in signal and image processing. In order to filter a
two dimensional image, it is best to convert the image to its frequency domain.
2D FFT is simply the FFT which has been applied to one direction followed by
the FFT implemented in another direction of the data. 2D FFT represents the
frequency spectrum in both dimensions, allowing filtering operations to be visually
studied.
To implement the 2D FFT, the following Equation 5.1 may be used:
F (u, v) =1
MN
M−1∑x=0
N−1∑y=0
f (x, y) e−j2π(uxM
+ vyN ) (5.1)
Similar to the 2D FFT, the inverse 2D FFT is simply Inverse FFT (IFFT) which
has been applied to both directions of the data. The Equation 5.2 represents the
2D IFFT:
f (x, y) =M−1∑u=0
N−1∑v=0
f (u, v) ej2π(uxM
+ vyN ) (5.2)
To better visualise the results obtained via implementing the 2D FFT, an
example is shown in Figure 5.6.
91
5. IMAGE PRE-PROCESSING
(a) Image (b) Magnitude Plot (c) Phase Plot
Figure 5.6: Implementing 2D FFT on a retinal image
Studying the magnitude plot of the obtained 2D FFT, reveals that most of
the energy is concentrated in the centre of the image. This corresponds with low
frequency data in the frequency domain, suggesting gradual changes in the image.
Moreover, in the result, there are no sharp lines away from the centre of FFT,
suggesting that there is no great energy in the higher frequencies.
The phase of the FFT is somewhat hard to interpret visually and generally looks
like noise. However, it holds a great deal of the information needed to reconstruct
the image. Therefore, including the phase plot in the results is essential as the
output of the research should not alter the original data and should have the
capability to reconstruct it if necessary.
The results obtained in this section are the preliminary stage of the processes
in the next Section and Section 7.1. Therefore, this process has been implemented
on over twenty different images and the results are included in Appendix D.
5.5 Sharpening the Retinal Image
There are times where certain features of the image need to be enhanced in order to
be detected. An example could be when the vessels in the retina are to be detected.
In such cases, it is advised to sharpen the image prior to feature localization. To
sharpen an image, the filtered image may be added to the original image. This
would result in highlighting the key features and emphasising on their edges.
In the study, to sharpen the image, the use of the 2D FFT and convoluting it
92
5. IMAGE PRE-PROCESSING
with a kernel of set value is suggested. It should be noted that for different image
sizes the kernel sizes may also vary based on the level of details required. For the
purpose of this study sSeveral different kernel sizes have been examined. In order
to observe the effect of the kernel, two kernels of 10×5 and 3×2 have been used
to compare their effects and the obtained results are displayed in Table 5.2. They
provided the best clarity in the results and observations of the studied images and
therefore were chosen to be implemented in the consecutive stages of the study as
well. The two kernels have been shown in Figure 5.7.
(a) 10×5 (b) 3×2
Figure 5.7: Used Kernels
Kernels are used to dilate the images. Dilation is a commutative process,
operating to grow or thicken and object in a binary image [144]. It can be used to
enhance a certain feature of the image. Opposite to dilation is erosion, in which
the object in the binary image shrinks or thins [144]. It may be used to remove
unwanted smaller objects, including the non-variable holes or dusts [144, 151]. It
should be noted that during both the erosion and dilation process, small cells,
noise and some details are lost, but the essential characteristics remains [151].
Comparing the outputs displayed in Table 5.2, it can be seen that the results
for the two suggested kernels slightly vary. The smaller kernel size, results in a
better sharpened image and so more details can be viewed in this case. On the
other hand, if larger details are of interest, the use of larger kernel size would be
more appropriate.
93
5. IMAGE PRE-PROCESSING
Table 5.2: Sharpening the retinal image using 2D FFT
10 × 5 kernel 3 × 2 kernel
Green Channel of Image
Magnitude Plot
Real Part of Spectrum
Imaginary Part of Spectrum
Filtered Image
Subtract Filtered Image from
Original Image
Inverse - Subtract Original Im-
age from Filtered Image
Moreover, the two kernel sizes which were selected and implemented were just
94
5. IMAGE PRE-PROCESSING
samples so that the effects of the variation in kernel size on the overall outcome
could be visualized. For different cases and capturing devices, this kernel size may
vary and so should be reset if necessary in order to provide the sharpest image
possible. This process has been implemented on all the images within the DRIVE
databas. The results are included in Appendix E and indicate similar outcomes
as the above discussion.
5.6 Trimming Regions
Many of the studies performed previously suggest that the outcomes from the
automated feature localisation stage are not 100% accurate. To overcome this
problem and enhance the results, in such cases the manual input from the user is
suggested to be used. The downfall of this would be that the outcome might vary
depending on the individuals with different experience levels. Moreover, in cases
where the expert opinion is not available the semi-automated system might not
provide the ideal result. In the study, for a fully automated detection process it
has been suggested to consider and resolve errors which result in the misdetection
of the feature of interest.
(a) (b)
Figure 5.8: Two examples of retinal fundus images. If observed closely, a brightfringe can be seen at the left hand corner of the image (b) which may result ininaccurate OD detection. The bright fringe cannot be seen in the image (a).
One of the common problems with obtained images is the unbalanced bright-
ness in the fringe of the rim, which is caused when the patients do not place their
95
5. IMAGE PRE-PROCESSING
eyes tightly against the capturing device. This leads to misdetection of the ROI,
which in this case is the OD.
In this study, 10% of the images with the database illustrated signs of fringe
noise presence.
To overcome this problem, study conducted by Zhang et.al. in 2010 [152]
has proved to be vital for system’s uniformities and accurate detection of ROI.
The authors introduced a pre-processing step, known as the fringe removal. They
proposed a trimming circle, where its center and radius were defined based on the
least-square fitting technique, previously suggested by Kasa [153].
The suggested trimming circle is represented by Equation 5.3, has its center
located at (Cx,Cy ) and its radius is shown in Equation 5.6. It should be noted that
in order to remove all the bright regions caused by ambient light, the estimated
radius is set to be smaller than the calculated radius [152].
X2 + Y 2 + (AX) + (BY ) + C = 0 (5.3)
Cx =−A2
(5.4)
Cy =−B2
(5.5)
r =
√A2 +B2
4− C(5.6)
Furthermore in this study, the OD region was considered to be the 0.5% of
the bright spots in the trimmed fundus image. The centroid of the region was
considered as the center. The ROI boundary was limited by considering a radius
twice those of the normal OD [152].
Implementing the circle on the image and then processing the fundus image to
detect the OD, resulted in successl rate of 96% detection. In the remaining cases,
the manual input of the user, was used to adjust the region of interest [152].
Since only 5% of the pixels has been considered as OD, this method may or may
not have the desired accuracy as different image intensity may reduce the precision
of localisation. Moreover, the OD was approximately determined by considering
96
5. IMAGE PRE-PROCESSING
a boundary twice hose of the normal OD. As a result, the chosen OD pixels may
not be correctly selected and so critical information may have been lost. Therefore
this study introduced a new methodology in localising the exact OD region, more
details have been provided in Chapter 8.
To further improve the accuracy of detection, in this study another circular
trimming region has also been suggested. Since not all fundus images appear to
be circular, the trimming region is further modified to provide the best possible
outcomes in such cases. More details are provided in section 5.6.2
5.6.1 Circular Trimming Region
As it can be seen the previously suggested technique was not ideal and there was
still a need for manual user input. To improve the results and the success rate of
detection, the previously proposed procedure has been re-examined and the new
approach suggested.
As it is known, a circle, centerd at (h, k) is represented by the equation:
(x− h)2 + (y − k)2 = r2 (5.7)
The result of expanding and rearranging this equation would be:
x2 + y2 − (2hx)− (2ky) + h2 + k2 = r2
x2 + y2 − (2hx)− (2ky) + h2 + k2 − r2 = 0 (5.8)
Comparing Equation 5.3 with that of Equation 5.8 suggests similarities between
the two, and hence equating them would provide:
X2 = x2 −→ X = x (5.9)
Y 2 = y2 −→ Y = y (5.10)
97
5. IMAGE PRE-PROCESSING
Now compare the third and fourth terms:
AX = −2hx
Based on Equation 5.9, it can be said that: X = x, therefore:
A = −2h −→ h =−A2
(5.11)
Moreover:
BY = −2ky
Based on Equation 5.10, it is known that: Y = y, therefore:
B = −2k −→ k =−B2
(5.12)
Comparing Equation 5.3 with the Equation 5.8 indicates that the constant
term is:
C = h2 + k2 − r2 (5.13)
Substituting Equations 5.11 and 5.12 into 5.13 and simplifying would result in:
C =
(−A2
)2
+
(−B2
)2
− r2
=A2
4+B2
4− r2 (5.14)
Using Equation 5.14, the variable r is made the subject:
r2 =A2 +B2
4− C
r =
√A2 +B2
4− C (5.15)
In Table 5.3, the suggested trimming region by Zhang et.al. and the suggested
trimming circle in the study is represented.
98
5. IMAGE PRE-PROCESSING
Comparing the two trimming regions, concludes that the location of the center
for both cases is the same and is represented by (h, k); however, as it can be seen
in Table 5.3 their radii are defined differently. Under the same conditions and
specifications, the calculated radius of the new proposed trimming technique is
smaller than those suggested by Zhang et.al.
It should be noted that similar to the previous technique, the radius of the
proposed trimming circle is also set to be smaller than the estimated radius. This
marginal variation in radius ensures that all the bright regions have been removed
from the image. The amount for the variation would depend on the number of
image pixels, general location of the OD in the image and its distance to the black
boundary.
Table 5.3: Comparison table of the proposed trimming circle with those suggestedpreviously in literature
Trimming region (Zhang et.al.) Proposed trimming circle
Equation X2 + Y 2 + AX +BY + C = 0 (x− h)2 + (y − k)2 = r2
CenterCx = −A
2
Cy = −B2
h = −A2−→ h = Cx
k = −B2−→ k = Cy
Radius r =√
A2+B2
4−C r =√
A2+B2
4− C
Therefore, when analyzing a new set of database, with different capturing set-
tings, it is suggested to visually observe a few of the retinal images so that if
required, the variation in radius could be changed for all the images within that
database.
99
5. IMAGE PRE-PROCESSING
5.6.1.1 Implementation
In order to implement the trimming region on wide range of databases and retinal
images with variety of resolutions and capturing settings, it is essential to create an
algorithm in which the required variables are detected for each individual image.
The key requirement in plotting the region is to locate its center. The steps
undertaken to determine the center of the analysed retinal images are as follow:
1. The first non-zero pixel is determined. The first pixel would be the one which
is not black and so is not part of the background black boundary. This would
represent the most left non-zero pixel in the retinal image.
2. Last non-zero pixel is then found. This would be the last pixel which is not
black and is located on the right hand side of the image.
3. To estimate the center, the horizontal and vertical pixel locations of the pin-
pointed pixels are used and the middle values are calculated and considered
as a preliminary location of the center.
4. Using the horizontal middle value found previously, the first and last non-zero
values in vertical directions are determined. These points would represent
the furthest top and bottom points where the pixel values are still non-zero.
5. Similarly, using the vertical middle value found in step 3, the first and last
non-zero values in horizontal directions are also determined.
6. Once the points are determined, their average values are taken, resulting in
the re-calculated center location of the trimming region. The final center
point can be seen as orange (+) sign on the images where the trimming
regions are plotted.
The other necessary value needed to plot the region, would be its radius. The
radius can easily be calculated using the difference between the number of pixels
from the center to any of the four previously founded points in the top, bottom,
left or right hand side.
100
5. IMAGE PRE-PROCESSING
Figure 5.9: Example of results obtained for plotting a trimming region. Thegreen (+) signs indicate the preliminary estimated points. The orange (+) signsindicate the calculated points, including the estimated center. The yellow circle isthe trimming region which has been plotted using the information.
Based on the above process and the obtained values for the radius and center
location, the trimming region may now be plotted as depicted in Figure 5.9. More-
over, since the above process is repeatable, it can be implemented on any given
RGB image, with any specifications.
5.6.1.2 Results and Discussion
The proposed circular trimming region has been implemented and results have
been displayed in this section. The first point to consider is to explain the im-
portance of using modified radius value instead of the estimated radius. Table 5.4
displays the results obtained from implementing the suggested trimming circle
using both estimated and modified radius values.
The left hand column represents the results obtained from implementing the
circular trimming method using the calculated radius, while the right hand column
shows the results for the same image, when a smaller radius has been used. This
small change in radius ensures that all the fringes are removed and the OD is
accurately detected.
The variation between the estimated radius and the implemented radius could
be defined based on the ROI. As mentioned previously, this type of error mainly
101
5. IMAGE PRE-PROCESSING
occurs when detecting the OD. Therefore, it is important to choose the radius so
that all the bright fringes are removed while the OD remains untouched.
Table 5.4: OD localisation using trimming circle
Trimming Region
(Estimated radius)
Trimming Region
(Modified radius)
Trimming Region
(Yellow)
Circularly Trimmed
Brightest Regions
(Possible OD)
Estimated OD
(Red)
102
5. IMAGE PRE-PROCESSING
Observing the results suggests that the use of proposed circular trimming re-
gion with modified radius would provide more accurate results for localization of
OD. However, not all fundus images are circular and so this method would be
insufficient and the results imprecise. Therefore, there is a need for another ap-
proach and so the author has suggested the use of elliptical trimming region when
the outcomes of the circular trimming is invalid.
5.6.2 Elliptical Trimming Region
As shown in Figure 5.10, depending on the setting of the capturing device, the
trimming region, may not be circular, and may be more of a truncated shaped.
Figure 5.10: Examples of retinal images using different capturing devices.
If the radius based on the short axis was calculated, the circular region may be
similar to those found in Figure 5.11. The obtained results clearly indicate that
some of the ROI has been cut out and hence the localisation will be inaccurate as
critical information has been removed.
(a) (b)
Figure 5.11: (a) Inaccurate circular trimming circle (yellow) for an elliptical shapedcaptured fundus image. (b) Trimmed image.
103
5. IMAGE PRE-PROCESSING
To avoid the loss of critical information, it is essential to revise the suggested
methodology. Hence, the long axis has to be considered and used to calculate
the radius of the estimated trimming circle or a secondary analysis has to be
implemented using an elliptical trimming region. Using the long axis to calculate
the radius of the circle would provide results as indicated in Figure 5.12.
(a) (b)
Figure 5.12: (a) Accurate circular trimming circle (yellow) for an elliptical shapedcaptured fundus image using long axis as the radius. (b) Trimmed image.
The results obtained in Figure 5.12 suggests that a circular trimming circle
may be sufficient to remove all the noise close to the black boundary. However, it
is also possible to use an elliptical trimming region as shown in Table 5.5.
Table 5.5: Proposed Circular and Elliptical Trimming Regions
Proposed Trimming Circle Proposed Trimming Ellipse
Equation (x− h)2 + (y − k)2 = r2 (x−h)2
a+ (y−k)2
b= 1
Center (h, k) (h, k)
Radius r =√
A2+B2
4−Cx−Radius = a
y −Radius = b
Referring to Table 5.5, it is apparent that the two regions only vary in radius,
104
5. IMAGE PRE-PROCESSING
where in an ellipse both short and long axis are taken into consideration, which is
basically what is assumed in the previous section for the circular trimming of an
oval shaped fundus image.
5.6.2.1 Implementation
The implementation would be similar to that of the circular trimming region,
with the minor variation of short and long axis. Table 5.6 shows examples of
the results obtained for both circular and elliptical trimming regions for different
fundus images.
Table 5.6: Implementation of both circular and elliptical trimming regions forcircular and elliptically shaped retinal fundus images
Circular Image Elliptical Image Elliptical Image
Trimming Region
(Yellow - Circular
Green - Elliptical)
Circularly
Trimmed
Elliptically
Trimmed
5.6.2.2 Results and Discussion
The obtained results suggests that for fundus images which appear to be circu-
lar, both the elliptical and circular trimming regions would approximately be the
same, hence the use of circular trimming region which has less variables would be
105
5. IMAGE PRE-PROCESSING
sufficient and provide good accuracy. However, for the truncated retinal images,
the two trimming regions of circular and elliptical will not be the same. In such
cases, the use of elliptical trimming region is suggested.
If the ROI is OD and the objective is to localise it, elliptical trimming region can
provide an accurate but faster results than those of the circular trimming region.
As discussed previously, this is due to the fact that the radius of the circular
trimming region may need to be re-calculated. However, if the elliptical region is
implemented, all the fringe noises are removed with the preliminary calculation of
the both short and long axis radii without the need for any recalculations. The
results can be seen in Table 5.7.
Table 5.7: OD Detection for Circular and Elliptical Trimming Region
Trimmed Image OD Detection
Circular
Trimming
Region
Circular Image
Elliptical Image
Elliptical
Trimming
Region
Circular Image
Elliptical Image
This is only true for cases where the OD or Macula localisation and their anal-
106
5. IMAGE PRE-PROCESSING
ysis are of importance. In other cases, such as vasculature detection, it is essential
to preserve data and information as much as possible. As a result, using ellipti-
cal trimming region, where great extent of data is removed, affects the accuracy
of detection significantly and is undesirable. In such cases, circular trimming re-
gion might be the better option. It should be noted that to improve the results
even further, multiple radius values may be used for individual images in order to
determine the best radius for the circular trimming circle.
In conclusion, using the suggested trimming circle and ellipse with adjusted
radius and applying them to the variety of data bases and re-examining the OD
detection using the proposed methodologies, suggested that the localisation is of
more accuracy and the detection rate is now 100% when fringe noise is present, in
comparison to the studies previously conducted in the literature.
There are also times when the whole image is too dark or too light. In such
cases, the detection of features become more difficult as the boundaries would be
less defined. To improve the detection precision in these circumstances, the image
intensity has to be adjusted by enhancing the image contrast. More details are
provided in the next section.
5.7 Contrast Enhancement
The previous section considered the effect of a localized variation in contrast and
how to overcome this problem using trimming regions. The overall results were
promising and the precision of the localization of ROI was improved.
However, the variation in contrast is not always confined to a specific region
of the image. The whole image may appear to be brighter or darker than the set
specification of the system and as a result the accuracy of detection is reduced for
the localization of ROI. In such cases another approach has to be taken.
Moreover, the flexibility of the automatic detection process is critical in iden-
tifying the ROI. There are occasions where the ROI detection has been affected
by the variation in contrast of the obtained images. In such cases, the proposed
system should still be able to proceed and perform its function successfully.
In majority of the studies performed previously, the grey scale image or the
green band of the coloured image was chosen for further processing. Similarly,
107
5. IMAGE PRE-PROCESSING
initially for this study, the green band was chosen to detect the features of interest.
However, there were times where the ROI detection proved to be inaccurate or
unresponsive. Visually revisiting and studying these images indicated that the
variation in contrast, might have been the cause.
Illumination in fundus retinal images is uneven and so the images may appear
to be brighter or darker. The variation in the lighting is dependent on the image
capturing system or the response of the retina itself. The non-uniform illumina-
tion adversely affects the ROI localisation precision and may even result in mis-
detection. In such cases it is essential to consider various contrast enhancement
techniques for re-adjusting the image colours. Once this stage is complete, the
remaining detection processes may be re-implemented and results obtained. Con-
trast enhancement is crucial in medical field as it can reveal information which
might have been otherwise missed or hidden from view.
There are two widely used approaches in contrast enhancement, the linear con-
trast stretching 1 or the histogram equalization 2. In the linear contrast stretching,
the dynamic range of the image is adjusted, while in the histogram equalization,
form the integral of the image histogram, the input and output relation is ob-
tained [154]. In this study, the most common approach in field of medicine is
chosen for further investigation which is the histogram equalization method.
From the available histogram equalization techniques, the Adaptive Histogram
Equalization (AHE) and the Adaptive Contrast Enhancement (ACE) are the most
popular [154].
The AHE algorithm uses the local histograms obtained from the gray values
of pixels. The image is separated into blocks. A particular pixel is enhanced by
interpolating its mapping function with its neighbouring four blocks [154].
The ACE method uses the unsharp masking technique in which the image is
separated into two masks using the low frequency filter. The high frequency mask
is obtained by subtracting the low frequency mask from the image. The amplified
1Linear contrast enhancement or linear contrast stretching is when the original values areexpanded into a wider range. As a result the subtle changes in variation become more apparent.
2Histogram Equalization is when both shape and distribution is taken under consideration.Each level in the displayed image has to have approximately equal number of pixels. This isachieved by stretching the regions with more pixels more than those with few pixels in thehistogram.
108
5. IMAGE PRE-PROCESSING
high frequency mask is then added to the original image in order to enhance the
image contrast [154].
Since in the histogram has been already obtained, the AHE technique which
uses the information obtained from a local histogram to map the gray values of
the pixels has been chosen and implemented.
Depending on the feature of interest, the approach undertaken for contrast
enhancement may vary. Therefore in the study a few different histogram equaliza-
tion methods have been implemented as the interest regions varied significantly in
characteristics.
5.7.1 New Necessary Step
To overcome the uneven illumination, the contrast of each of the images is to be
modified. For each of the contrast enhancement methods, the histogram of the
original image is changed and adjusted to form a new histogram known as the
Desired Histogram (DH).
5.7.1.1 Intensity Adjusted
The first approach has been to modify the intensity variance of the image, such
that 1% of the low and high intensities of the gray scaled image is saturated.
(a) Original Histogram (b) Intensity Adjusted
Figure 5.13: Example of the effect of Intensity Adjustment
109
5. IMAGE PRE-PROCESSING
Figure 5.13 illustrated an example of the changes which occur to the histogram
when its intensity is adjusted. This may be useful, in particular when the original
image is quite dark and so the affect will increase the contrast of the overall image.
However, when noise is present in either of the low or high intensity bands, this
method may not be reliable as it magnifies the error.
5.7.1.2 Histogram Equalization
The second approach is the Histogram Equalization (HE) method [144, 155]. It
basically involves modifying and equalizing the intensity of each image so that the
illumination effects have been minimized. An example of the effect of HE method
on a histogram can be viewed in Figure 5.14.
(a) Original Histogram (b) HE
Figure 5.14: Example of the effect of Histogram Equalization
In this section, a flat DH is formed and applied to the image [144]. It is as
follow:
Desired Histogram =ones(1, n) ∗ Pdt(Size(A))
n(5.16)
Where Pdt is the product of array element in A.
The DH ensures that the grey scale transformation T is minimised by:
|c1(T (k))− c0(k)| (5.17)
110
5. IMAGE PRE-PROCESSING
In the Equation 5.17 the c0 is the cumulative histogram of A and the c1 is the
cumulative sum of DH for all intensities of k.
It should be noted that this equation is constrained such that T must be
monotonic. Moreover, c1 (T (a)) should not overshoot c0 (a) by more than half
the distance between the histogram counts at a.
To map the grey level into their new values, the DH uses the b = T (a) trans-
formation.
5.7.1.3 Adaptive Histogram Equalization
The final approach which is in most cases a more effective method than the HE,
is the Adaptive Histogram Equalization (AHE) method. It is more commonly
known as the Contrast Limited Adaptive Histogram Equalization (CLAHE) as it
concentrates on a small region of the image. It follows the work performed by
Zuiderveld [156].
Figure 5.15 shows the effect of CLAHE when it is applied on a sample his-
togram.
(a) Original Histogram (b) CLAHE
Figure 5.15: Example of the effect of Contrast Limited Adaptive Histogram Equal-ization
CLAHE separates the image into smaller regions and works in enhancing the
contrast in each of those sections, therefore the histogram output approximately
111
5. IMAGE PRE-PROCESSING
is the same as the specified DH. In order to remove the boundaries formed by each
region, the bilinear interpolation can then be applied to smooth out the output
image.
This technique is quite useful as it minimizes the amplification of the noise
present. This is due to the fact that the image is separated into smaller regions
and each section is analyzed separately, reducing the effect of the noise to the
surrounding regions. However, due to computational difficulties, this process may
also take longer.
5.7.2 Results and Discussion
In this section, the effect of each of the different contrast enhancement methods
discussed above is visualised. A sample image has been selected to represent the
effect of each method and how enhancing the contrast of the image may help in
accurate localisation of the ROI.
The outcome is clearly depicted in Table 5.8. As it can be seen, previously
while detecting the OD as the ROI using the green band of the image; the result
was inaccurate. However, after implementation of the three approaches, the OD
was correctly localised.
The variation in contrast and the effect of each of the techniques is apparent
in the image. As mentioned earlier, depending on the area of the interest, the im-
plemented methodology can then be chosen. For example for the OD localisation,
it is better if the image is not too bright since the OD is the brightest region in
the retinal image. If the image is too bright, there is a possibility of misdetec-
tion. However vasculatures are best visible in high contrast and bright images.
Therefore depending on the feature of interest, the chosen methodology to adjust
contrast could vary.
For the purpose of this study, the discussed methodologies proved to be ad-
equate and the implemented approaches tend to provide sufficient information.
For different applications, one or two of the methodologies have been useful. For
example, in vasculature detection the HE method proved to reveal more intricate
details, while the AHE method was used primarily for OD localisation.
112
5. IMAGE PRE-PROCESSING
Table 5.8: OD localization for contrast enhanced images.
Image OD localisation
Green Channel of
Image
Intensity adjusted
image
Histogram
equalised image
Adaptive histogram
equalised image
113
5. IMAGE PRE-PROCESSING
5.8 Summary
In this chapter the pre-processing stage of image processing was applied, so that the
overall accuracy of feature detection could be enhanced for automated processes
with minimal user input.
The common processes including the conversion of the images from colour to
grey scaled and green band images, masking, filtering and image sharpening were
implemented.
Based on the previous literature, the coloured images are best to be converted
into the grey scaled images or to their primary components. This enhances the
processing time, while preserving the image details. The study agreed with the
literature that green channel of the image provided the best level of details for the
purpose of this research and was selected to be used for the consecutive steps of
the image processing process.
To define the ROI, a new thresholding procedure was suggested to mask the
images. This new procedure automatically locates the location of the first major
minimum in the image, separating the ROI from the black background. The
obtained results are promising, demonstrating rapid but exact localisation of the
ROI for masking the retinal images. The accuracy is very similar to the Otsu
method; however, it is computationally less complicated and faster.d
The use of 2D FFT filter was suggested also suggested in this study as a
software filter in order to improve the processing time in the consecutive stages of
the image processing. Using the 2D FFT in conjunction with a kernel was then
used to sharpen the image so that the key features of interest in the fundus image
could be enhanced. The smaller kernel size proved to provide more details about
the image while the larger kernel size displayed the overall outlay of the features.
This characteristic has been used in the coming chapters for localisation of vessels
in the retinal images.
Furthermore, two main factors which may lead to imprecise localization or
misdetection of ROI were considered, including the presence of fringe noise or
localization of a desired feature in low contrast retinal images.
The fringe noise mainly occurs during the capturing where the ambient light
affects the image when the patients’ eye is not placed directly in front of the device.
114
5. IMAGE PRE-PROCESSING
The presence of this bright spot may result in misdetection of the ROI, in particular
when localizing the OD. To eliminate this error, the use of new trimming regions
was suggested so that depending on the shape of the fundus image, the illuminated
noise could be removed automatically. The results for both circular and elliptical
trimming regions proved to be promising, with improvements in overall rate of OD
detection to 100% in comparison to the previously conducted studies.
The other factor which was considered in this chapter was to implement con-
trast enhancement methodologies, so that the ROI could be more easily distin-
guished and localised. Majority of the researches tend to not perform this step
and only use the grey scaled or green band of the colours image in the analysis.
However, for an automated system, it was observed that enhancing the image con-
trast can play a significant role in localisation of the feature. Different histogram
equalization approaches were considered and implemented, including the Intensity
Adjustment, Histogram Equalization and Adaptive Histogram Equalization. The
precision for feature detection has improved as a result, especially when Histogram
Equalization was used in vascular detection and Adaptive Histogram Equalization
was used for OD localisation.
In conclusion of the chapter it can be said that the pre-processing stage im-
proved the outcome of the detection process and increased its success rate. It also
reduced the amount of manual user input needed for feature localization system
of retinal images.
115
CHAPTER 6
IRIS AND PUPIL LOCALISATION AND EXTRACTION
Introduction Literature Review Thesis Outline Conclusion
Image Acquisition Image Pre-Processing
Feature Localisation:
—Iris
— Pupil
Feature Extraction:
— Center
— Area
Figure 6.1: Chapter Six Outline of Image Processing Stages
6.1 Overview
The next main step in image processing after image acquisition and pre-processing
is feature localisation and extraction. In the literature review chapter, some of the
key features of the eye have been identified to be important in many applications
of ophthalmology and biometrics. In this chapter, Iris and Pupil of the eye have
been considered as the features of interest.
Recent studies and applications of biometrics authentication which relates to
the human characterisation and identification, suggests that uniqueness of individ-
ual Iris pattern can be used to separate and distinguish people with extremely high
116
6. IRIS AND PUPIL LOCALISATION AND EXTRACTION
accuracy. Hence, many methodologies have been introduced and implemented in
order to localise the Iris.
Accurate detection of the Pupil boundary can also play a significant role in
the field of ophthalmology as well as biometrics. Accurate detection of the Pupil
and Iris boundary specify the exact Iris region which can then be used for exact
pattern extraction as well as disease diagnosis, treatment and monitoring stages.
For example, during a treatment procedure, such as the case of cataract surgery,
detection of the marginal variations of the size of Pupil boundary may minimise the
occurrences of complications to a great extent. Hence it is important to accurately
detect the Pupil boundary and its changes.
Based on the above factors, new methodologies for fast localisation of the Iris
and Pupil have been proposed in this chapter. Moreover, approaches have been
suggested to automatically detect and measure the center and area of the features
so that medical practitioners could use this information to identify changes due to
disease or complication.
The two essential steps in image processing, the feature localisation and feature
extraction for Pupil and Iris are considered in detail and different approaches are
suggested and applied for the betterment of the final outcomes.
6.2 New Technique for Iris/Pupil Localisation
Exact boundary detection of the Iris and Pupil restricts the affected area and fur-
ther analysis may provide the ophthalmologist with more insight to the severity of
the disease. The information may aid ophthalmologists with all level of experience
to better diagnose and treat the patients.
As discussed in section 3.2.2.1, many different approaches have been considering
Iris and Pupil detection, each having their own advantages and disadvantages.
Amalgamates process has been suggested to be used in this section. This is to
ensure that the actual desired region is selected and more accurately detected. In
order to do so, the results of two different techniques have been fused together,
creating a single mask which segments the ROI, which is then applied to the
original data to define the ROI. In this case the Iris and Pupil have been detected
using this methodology.
117
6. IRIS AND PUPIL LOCALISATION AND EXTRACTION
Figure 6.2 illustrates the flowchart of the overall procedure of the proposed
technique by the author for detection of Iris and Pupil [9, 12].
EyeImage
Acquisition
Image
Pre-Processing
Iris/Pupil
Localisation
Method 1 Method 2
Mask1 Mask2
Overall Mask
(Mask1 + Mask2)
Matching
(with previous data)
Interpretation
(Feature Extraction)Display
Complication?
Alarming System
Revise
Procedure
Continue
NoYes
Figure 6.2: Proposed steps for Iris and Pupil localisation.
Normally any single approach may have its own advantages and disadvantages,
affecting the overall outcome. To check the validity of the result and verifying that
the detected region is in fact the desired ROI, it is best to double check the outcome
with another methodology as well.
From the studied literature, two methodologies have been chosen and imple-
mented to investigate and prove this concept. The thresholding approach sug-
gested by Masek [107] and the active contouring procedure introduced by Rit-
ter [108] have been implemented following the suggested. Detailed explanation of
the two processes and their advantages were included in Section 3.2.2.1.
Figure 6.3 is an example of the possible results which might be obtained from
two different techniques. The results for Iris localisation from technique one is in
green and the technique two is in red. It can be seen that each of the techniques are
not ideal and have missed some critical information. To overcome this problem
and ensure that none of the required information is removed, the best possible
solution would be to define the ROI as the combination of the regions by both
techniques.
118
6. IRIS AND PUPIL LOCALISATION AND EXTRACTION
Figure 6.3: Example of the possible inaccurate results obtained from two differentIris localisation techniques. Results from approach one and two are outlined ingreen and red respectively.
Usually the undertaken approaches are similar in outcome, with minor varia-
tions. Therefore creating an overall mask, combining the two approaches reduces
the detection error and localizes the region with more precision.
Moreover, with an increase number of different approaches, the computational
time also increases, therefore in this study the results from two approaches have
been chosen to be combined. In cases where time is not of an essence, the results
from multiple techniques maybe combined for higher precision.
Another point to consider is that this procedure should be designed such that it
could be applied in the treatment stages of ophthalmology. At this stage, since the
chosen images were from open source databases the main objective has been the
localisation and extraction in a timely manner. Therefore, there is an assumption
that no eyelashes and eyelids can be viewed in the images and so their removal has
not been taken under consideration. This agrees with the treatment procedures
were the eyes are clamed open. As a result of this assumption the computational
complexity has been reduced significantly as the unwanted noise is not present
Additionally, the images for the investigation have been chosen such that they
were clear and not blurred as a result of the slight movements of the eye and the
head. Therefore filtering for deblurring was not considered further in the study.
Furthermore, an exact localisation of Iris and Pupil boundary is of interest in
this case and so no assumptions have been made in regards to their shape being
circular or elliptical. Therefore, the two chosen methodologies from the literature
would need to exactly detect the boundaries without approximating them.
119
6. IRIS AND PUPIL LOCALISATION AND EXTRACTION
It should be noted that Iris and Pupil are similar in shape, so the same method-
ology with different parameters may be used to detect these features. This reduces
the complication of the implemented algorithm and so is more feasible to be used
in an automated system.
6.3 Implementation
In this section, the proposed methodology has been implemented and the results
are observed. An example, using the original image shown in Figure 6.4, has been
used for better representation of the possible outcomes.
Figure 6.4: Original image used for localisation of Iris and Pupil
The results are promising and the feature of interest has been accurately located
in comparison to the results of each of the techniques separately. Similar results
have been observed when detecting the Pupil. Example of the result obtianed
when localising the Iris and Pupil using the proposed methodology can be viewed
in Figure 6.5.
(a) Localised Iris (b) Localised Pupil
Figure 6.5: Result obtained when localising the iris and pupil outer boundariesusing the proposed new algorithm
120
6. IRIS AND PUPIL LOCALISATION AND EXTRACTION
Moreover, the step by step results for Iris localisation using the proposed
methodology has been shown in Table 6.1.
Table 6.1: Example of Iris Localisation Results
Method 1 Method 2
Detection
Noise Removal
Mask
Feature Localisation
It should also be noted that majority of the pre-processing steps are the same
for different procedures, hence it is only the last stage of Iris and Pupil localisation
which may vary between the procedures. As a result the overall computational
time varies mainly due to localisation stage.
Since the two methodologies are being performed concurrently, the processing
time is also reduced in comparison to if the procedures were to be performed
separately and that is a desirable outcome for an automated process. For the
chosen methodologies the overall processing time was about 2-5 seconds.
The suggested process has been performed on over twenty different eye images.
Six samples of the obtained results for Iris localisation are displayed below in
Table 6.2.
121
6. IRIS AND PUPIL LOCALISATION AND EXTRACTION
Table 6.2: Iris localisation for different images.
Original Image Overall Mask Iris Localisation
Image 1
Image 2
Image 3
Image 4
Image 5
Image 6
The results are consistent and show reasonable robustness for detection of
the Iris boundary. In cases where the pigmentation of the Iris is lighter some
misdetection is observed, such as the case in the bottom left hand corner of Image
2 in Table 6.2. This is mainly due to the error in thresholding approach where there
is less contrast between the Iris and Sclera. To overcome this problem, it is best
122
6. IRIS AND PUPIL LOCALISATION AND EXTRACTION
to enhance the contrast using the suggested approach in Section 5.7. This ensures
that the contrast between the two regions are maximised and so the accuracy of
detection is improved.
6.4 Iris and Pupil Extraction
Once the Iris and Pupil have been detected and the boundaries have been localised,
the next step is to extract the feature information. The general information needed
are the center location and the area of the ROI. As a result in this section some
approaches have been suggested to obtain this information.
6.4.1 Center Decection
A simple approach has been used to estimate the location of the center of the Iris
and Macula based on the detected boundary of the localised region. To do so, the
following steps have to be undertaken:
1. Mask the ROI, so that the desired region is represented by ”1” and all sur-
rounding region are set as background and have ”0” pixel value.
2. First non-zero pixel is determined. This pixel would not be black and so
is part of the estimated Iris or Macula. It would represent the most left
non-zero pixel in the image.
3. Last non-zero pixel is then determined. This would be the last pixel of the
ROI and is located on the right hand side of the image.
4. In order to estimate the center of ROI, the horizontal and vertical pixel
locations of the first and last non-zero pixels are used to calculate the middle
point which may be considered as a preliminary location of the Iris and or
Pupil center.
5. To improve the accuracy of center estimation, the horizontal middle value
found in the previous step can then be used to determine the first and last
non-zero values in vertical directions. These points would represent the fur-
thest top and bottom points where the pixel values are still non-zero.
123
6. IRIS AND PUPIL LOCALISATION AND EXTRACTION
6. Similarly using the vertical middle value found in step 4, the first and last
non-zero values in horizontal directions can be determined.
7. Once the points have been determined, their average values are taken, re-
sulting in the re-calculated center location of the Iris and Pupil. An example
of the possible results from this process, will be shown in the OD localisation
Section 8.4.1, Figure 8.6. The final center point of the ROI can be seen as
blue (+) sign on the image.
6.4.2 Area Calculation
Once the center has been localized, the next step is to calculate the area of the
ROI, in this case the Iris and Pupil. Three suggested approaches are as follow:
The first approach could be to approximate the ROI as being circular, and
use the radius (R) to detect the area (A). The radius can be calculated using the
distance between the estimated center and the four non-zero pixels found in the
suggested center localisation approach.
Once the radius has been defined, the area can be calculated using:
A = πR2 (6.1)
The second approach could be to use the perimeter (P) of the ROI to determine
its area. The perimeter can be found more accurately by considering the ROI
pixels. In a binary image of the ROI, the pixels are considered to be part of the
desired region, if they are non-zero and are connected to at least one other non-zero
pixel.
Once the perimeter is calculated, the area of the ROI can then be estimated,
since:
P = 2πR (6.2)
Therefore:
R =P
2π
124
6. IRIS AND PUPIL LOCALISATION AND EXTRACTION
Substitute R back into Equation 6.2:
A = π
(P
2π
)2
=πP 2
4π2
=P 2
4π(6.3)
Using the Equation 6.3 the area of the ROI can be calculated.
Another approach used to determine the area of the ROI is to determine the
total number of the non-zero pixels. This value represents the area of the ROI
since the value of the desired region is set to ”1”, while its surrounding has ”0”
pixel value.
6.5 Summary
Iris and Pupil of the eye have been localised and their key information such as
center and area have been detected. Due to similarities in shape of both Iris and
Pupil, the proposed procedures to localise and extract these features were the
same.
In this study, the exact boundary detection, simplicity of the procedure and
the speed of detection were of interest, hence the proposed approach and the
methodologies were chosen accordingly.
Since each procedure has its own advantage and disadvantage, a new procedure
was proposed which was to obtain results from two different approaches and then
combine the outcomes to create a single mask covering both regions. The mask
could then be used to detect the ROI which in this case were the Iris and Pupil.
In this case, the thresholding and active contouring methods were selected and
their results were combined to create the mask for the ROI. Moreover, since both
methodologies were performed at the same time, the overall processing time is not
increased significantly. The obtained results for Iris and Pupil localisation proved
to be more precise, with less loss of critical information and with a reasonable
125
6. IRIS AND PUPIL LOCALISATION AND EXTRACTION
computational timing.
Once the Iris and Pupil were detected, their center, radius and area were then
calculated. The location of the center was approximated by finding the middle
value in horizontal and vertical direction within the detected boundary. The ra-
dius was then calculated by measuring the distance between the center and the
boundary. Using the equation for the area of the center, the area of Iris and
Pupil were approximately calculated. For a more accurate area, the equation for
perimeter of the circle was used to calculate the area.
The outcomes have been beneficial for the fields of biometrics and optometry
as the Iris and Pupil were successfully detected and their important information
extracted.
126
CHAPTER 7
RETINAL VESSELS LOCALISATION AND
EXTRACTION
Introduction Literature Review Thesis Outline Conclusion
Table 7.1: Modeling and implementation of different filters for vessel detection
Filter 10 × 5 kernel 3 × 2 kernel
Laplacian Filter
Prewitt Filter
Circular Average Filter
Average Filter
Median Filter
132
7. RETINAL VESSELS LOCALISATION AND EXTRACTION
Table 7.1: Modeling and implementation of different filters for vessel detection
Filter 10 × 5 kernel 3 × 2 kernel
Weiner Filter
Gaussian Filter
The results obtained from Canny filters appear to be more noisy than the
desired results. This is mainly due to the chosen thresholds. Since each image
would require its own specific threshold values, this method may not be feasible
and desirable for the purpose of this study.
Results from Sobel, Prewitt and Laplacian filters appear to be very similar.
The techniques consider variation in gradient and since the background gradient
is considered in the process, the results do not have the required clarity. The
remaining processes provided similar responses, much clearer than the Sobel, Pre-
witt, Canny and the Laplacian operations.
The results from the Circular Average filter and Average filter was very similar
due to the similarities in the process. However, under similar conditions the Av-
erage filter provided more vasculature details. Comparing the Average filter with
Median filter revealed that median values may result in more noise detection.
Based on observations, it can be said that the Average filter, Median filter,
Weiner filter and Gaussian filter revealed more details and clarity for vascular
detection.
To improve the results even further, several combinations of these filters were
133
7. RETINAL VESSELS LOCALISATION AND EXTRACTION
also examined and the results have been displayed in Table 7.2.
Table 7.2: Combining results of different filters
Filter 10 × 5 kernel 3 × 2 kernel
Laplacian of Gaussian Filter
Addition of Weiner and Me-
dian Filter
Average Filter of Weiner Fil-
tered image
Weiner Filter of Median Fil-
tered image
Gaussian Filter of the
Weiner Filtered image
134
7. RETINAL VESSELS LOCALISATION AND EXTRACTION
The results indicate that larger size kernel windows reveal more detail but
have more noise and unwanted error as well. In these cases the accuracy of the
procedures reduces, while the processing time increases. Therefore, for a general
overview of the vessels smaller size windows are preferred in this case.
Moreover, by combining and implementing multiple filters, the results appear
to have improved the clarity of the vasculature detection. This is quite apparent
in three cases of when an Average filter or Gaussian filter was applied to a Weiner
filtered image or when the Weiner filter was applied to the Median filtered image.
All the three cases revealed similar results, however in the case of applying Weiner
filter to the Median filtered image; it appears that some more minor details can
be viewed.
This is due to more emphasis of the locations of the vessels. In the original
results from Table 7.1, the best results for vessels localisation was found by imple-
menting a single filter to sharpen or blur the images. In this case, the emphasised
vessels are further highlighted by combining the filters and noise removal. In the
preliminary analysis, Median filter revealed more detailed structure of the vessels
while Weiner filter which is commonly used for noise removal highlighted the ves-
sels more clearly. Therefore, applying the Weiner filter to the obtained results
of the Median filter would have removed the unwanted noise and revealed more
vasculature structures. This agrees with the observations and the findings of this
study.
For confirmation of the observation, this process has been implemented on
over twenty different images. The average processing time for the retinal vessel
localisation was about 12-15 seconds. The results for five of the images has been
displayed in the following Tables 7.3 to 7.7.
135
7. RETINAL VESSELS LOCALISATION AND EXTRACTION
Results for Image 1:
Table 7.3: Vessel localisation for Image 1
Filter 10 × 5 kernel 3 × 2 kernel
Original Image - 2D FFT
Filtered Image
Sobel Filter
Canny Filter
Laplacian Filter
Prewitt Horizontal Edge
Emphasizing Filter
136
7. RETINAL VESSELS LOCALISATION AND EXTRACTION
Filter 10 × 5 kernel 3 × 2 kernel
Circular Average Filter
Average Filter
Median Filter
Weiner Filter
Gaussian Filter
137
7. RETINAL VESSELS LOCALISATION AND EXTRACTION
Filter 10 × 5 kernel 3 × 2 kernel
Laplacian of Gaussian Filter
Addition of Weiner and Me-
dian Filter
Average Filter of Weiner Fil-
tered image
Weiner Filter of Median Fil-
tered image
Gaussian Filter of the
Weiner Filtered image
138
7. RETINAL VESSELS LOCALISATION AND EXTRACTION
Results for Image 2:
Table 7.4: Vessel localisation for Image 2
Filter 10 × 5 kernel 3 × 2 kernel
Original Image - 2D FFT
Filtered Image
Sobel Filter
Canny Filter
Laplacian Filter
Prewitt Horizontal Edge
Emphasizing Filter
139
7. RETINAL VESSELS LOCALISATION AND EXTRACTION
Filter 10 × 5 kernel 3 × 2 kernel
Circular Average Filter
Average Filter
Median Filter
Weiner Filter
Gaussian Filter
140
7. RETINAL VESSELS LOCALISATION AND EXTRACTION
Filter 10 × 5 kernel 3 × 2 kernel
Laplacian of Gaussian Filter
Addition of Weiner and Me-
dian Filter
Average Filter of Weiner Fil-
tered image
Weiner Filter of Median Fil-
tered image
Gaussian Filter of the
Weiner Filtered image
141
7. RETINAL VESSELS LOCALISATION AND EXTRACTION
Results for Image 3:
Table 7.5: Vessel localisation for Image 3
Filter 10 × 5 kernel 3 × 2 kernel
Original Image - 2D FFT
Filtered Image
Sobel Filter
Canny Filter
Laplacian Filter
Prewitt Horizontal Edge
Emphasizing Filter
142
7. RETINAL VESSELS LOCALISATION AND EXTRACTION
Filter 10 × 5 kernel 3 × 2 kernel
Circular Average Filter
Average Filter
Median Filter
Weiner Filter
Gaussian Filter
143
7. RETINAL VESSELS LOCALISATION AND EXTRACTION
Filter 10 × 5 kernel 3 × 2 kernel
Laplacian of Gaussian Filter
Addition of Weiner and Me-
dian Filter
Average Filter of Weiner Fil-
tered image
Weiner Filter of Median Fil-
tered image
Gaussian Filter of the
Weiner Filtered image
144
7. RETINAL VESSELS LOCALISATION AND EXTRACTION
Results for Image 4:
Table 7.6: Vessel localisation for Image 4
Filter 10 × 5 kernel 3 × 2 kernel
Original Image - 2D FFT
Filtered Image
Sobel Filter
Canny Filter
Laplacian Filter
Prewitt Horizontal Edge
Emphasizing Filter
145
7. RETINAL VESSELS LOCALISATION AND EXTRACTION
Filter 10 × 5 kernel 3 × 2 kernel
Circular Average Filter
Average Filter
Median Filter
Weiner Filter
Gaussian Filter
146
7. RETINAL VESSELS LOCALISATION AND EXTRACTION
Filter 10 × 5 kernel 3 × 2 kernel
Laplacian of Gaussian Filter
Addition of Weiner and Me-
dian Filter
Average Filter of Weiner Fil-
tered image
Weiner Filter of Median Fil-
tered image
Gaussian Filter of the
Weiner Filtered image
147
7. RETINAL VESSELS LOCALISATION AND EXTRACTION
Results for Image 5:
Table 7.7: Vessel localisation for Image 5
Filter 10 × 5 kernel 3 × 2 kernel
Original Image - 2D FFT
Filtered Image
Sobel Filter
Canny Filter
Laplacian Filter
Prewitt Horizontal Edge
Emphasizing Filter
148
7. RETINAL VESSELS LOCALISATION AND EXTRACTION
Filter 10 × 5 kernel 3 × 2 kernel
Circular Average Filter
Average Filter
Median Filter
Weiner Filter
Gaussian Filter
149
7. RETINAL VESSELS LOCALISATION AND EXTRACTION
Filter 10 × 5 kernel 3 × 2 kernel
Laplacian of Gaussian Filter
Addition of Weiner and Me-
dian Filter
Average Filter of Weiner Fil-
tered image
Weiner Filter of Median Fil-
tered image
Gaussian Filter of the
Weiner Filtered image
Studying the results, suggest that the preliminary observations were correct.
Once again the results indicated that the smaller kernel size reduced the noise,
150
7. RETINAL VESSELS LOCALISATION AND EXTRACTION
while the larger kernel size showed more details. Since the noise would cause
confusion and may lead to misjudgement, the results for smaller kernel size are
preferred.
From the applied filters, the Average filter, Median filter, Weiner filter and
Gaussian filter revealed more information and displayed a better and clearer results
in comparison to the other applied filters.
Similarly, the combination of the filters once again showed that the best local-
isation of the vessels were obtained for the cases were the noise filtering Weiner
filter was applied to the Median filtered images.
7.4 Retinal Vasculature Extraction
Several key features of the vessels including its turosity and variation in diameter
have been considered in details in the literature. However, the disease which was
considered in this case was the ROP. As mentioned in Section 2.4 , ROP occurs in
premature infants and the main distinguishable feature of this disease is that the
vessels are affected as they are not well developed. The only cure for an irreversible
blindness in these infants is to apply laser treatment to the end point of affected
vessels. Therefore, in this case detecting end-point of vessels was investigated
further.
As a result, in this section, an approach has been suggested for localising the
end point of the vessels.
7.4.1 Localisation of the End Point of Vessels
In this section a method has been suggested in order to localise the end point of
the vessels.
1. Localise the retinal vessels from the fundus image.
2. Mask the vessels, so that they have the pixel value of ”1” and the remaining
background areas of the retina have the ”0” pixel value.
3. Trace the location of the vessels. If necessary burst or shrink the vessels to
the desired thickness.
151
7. RETINAL VESSELS LOCALISATION AND EXTRACTION
4. Using template matching, the end point of the vessels can then be calculated.
By considering a window of 3×3, where the desired non-zero pixel is located
at the center of it, the end point can then be determined, if the middle pixel
is only surrounded by one other non-zero pixel. Else, the desired pixel may
have been located in the middle of the vessel. Figure 7.3 shows an example
of a vessels end point.
Figure 7.3: Some of the possible vessels end point using template matching
7.5 Summary
One of the most important key features of retina is the vasculature. The aim of
this chapter was to introduce a new method to localise the retinal vessels and
determine their end-points.
To reduce the processing time of the localisation procedure, the edges of the
vessels have been detected by applying multiple different filters to the 2D FFT
image which was prepared in Chapter 5. From the studied edge detection filters,
the Average and Gaussian filters applied to Weiner filter and the Weiner filter
applied to the Median filter provided the best possible vessel detection. The vessels
in these cases were clearly visible and more easily distinguishable in comparison
to the other filters and the original images. However, the best visible result was
for the case were the Weiner filter was applied to the Median filtered image.
The extracted feature of the vessels was detecting their end-points. This may
be useful in studying the vasculature growth throughout the retina and diagnosing
diseases such as ROP. Furthermore, it may also aid the ophthalmologists in treating
such diseases as the areas at risk would be highlighted. To achieve this, template
matching was applied to locate the last non-zero pixels. These pixels would only
152
7. RETINAL VESSELS LOCALISATION AND EXTRACTION
have one other neighbouring non-zero-pixel and hence can be defined as the end
point of the vessels.
Successful, localisation and extraction of the vessels and their end-point were
the outcomes of this research. The short processing time of only a few seconds,
allow this process to be used in many diagnostic tools and a guide for ophthalmol-
ogists.
153
CHAPTER 8
OPTIC DISK AND MACULA LOCALISATION AND
EXTRACTION
Introduction Literature Review Thesis Outline Conclusion
Image Acquisition Image Pre-Processing
Feature Localisation:
—Optic Disk
— Macula
Feature Extraction:
— Center and Radius
— Area
— Cup to Disk ratio
Figure 8.1: Chapter Eight Outline of Image Processing Stages
8.1 Overview
In the literature review chapter, some of the key features of the eye have been
identified to be important in many applications of ophthalmology, including the
OD and Macula of the eye.
Diseases such as Glaucoma are detected using the OD, which is the brightest
region in the retinal image. Glaucoma is the second leading cause of irreversible
visual loss and blindness. Hence, to minimise vision loss in patients, early detection
and treatment of Glaucoma is crucial.
154
8. OPTIC DISK AND MACULA LOCALISATION AND EXTRACTION
Ophthalmologists diagnose Glaucoma by observing the visible changes which
occur at the OD. The diameter of the OD can be used as the preliminary in-
dication of the susceptibility of the patient to Glaucoma. As a result, the OD
localisation and measurement of its area are other key features of interest in the
field of ophthalmology.
Another main feature in retinal images is the macula, which is approximately a
dark circular region in the images. Macula may also help experts in their prognosis
and so its detection is important.
8.2 New Technique for Optic Disk Localisation
Over the past years, many methods have been suggested for detection of OD, each
having their own benefits and restrictions. Some of these approaches resulted in
localising OD center while others estimating its boundaries. One of such methods
has been the thresholding technique.
Thresholding technique has been widely used in the past to detect different
features of a retinal image. In this study, the thresholding approach has been used
to approximate the location of the OD. The reason being is that this approach
would provide an exact boundary of the OD in comparison to majority of the
other available techniques which assume OD to have a circular or an elliptical
shape.
EyeImage
Acquisition
Image
Pre-Processing
Optic Disk
Localisation
Histogram of ROI
Apply Threshold
/Brightest Pixels
Define OD Region
Mask OD Region
Interpretation
(Feature Extraction)
Center
Radius
Area
Cup to Disk Ratio
Display
Figure 8.2: Proposed steps for Optic Disk localisation.
155
8. OPTIC DISK AND MACULA LOCALISATION AND EXTRACTION
As mentioned previously, OD appears as a circular bright spot on the fundus
retinal image. Using this characteristic, the adaptive thresholding technique has
been implemented to detect the brightest pixels of the image. Figure 8.2 depicts
the flowchart of the overall procedure of the proposed technique for detection of
Optic Disk.
In the first step, the bright pixels have been detected and the outcome has
then been displayed as a binary image. This has been achieved using automated
adaptive thresholding for each individual image. Figure 8.3, displays a sample
histogram of this process and the set threshold.
The result is somewhat noisy; therefore filters have been applied to remove this
noise. Noise removal of the result by median filtering has proven to be successful.
Figure 8.3: The gradient plot histogram used to set the threshold for the ODlocalisation.
As demonstrated in Figures 8.3 and 8.4, the OD region is defined by determin-
ing the pixels with the higher intensity values. Since not all the images have the
same intensity and brightness, the threshold has to be set individually. Defining
the threshold is easier in cases like Figures 8.4a and 8.4b as the majority of the
bright pixels are bundled together and easier distinguishable in the Gradient Plot.
However, there are times where defining this region would be more difficult such
as the case in Figure 8.4c. In such cases, the threshold has been set as the first
156
8. OPTIC DISK AND MACULA LOCALISATION AND EXTRACTION
occasion where a few closely located pixels have intensity values greater than 200
gray levels.
(a) Sample 1
(b) Sample 2
(c) Sample 3
Figure 8.4: Example gradient plot histograms and set thresholds for OD localisa-tion for different images.
At this point, a reasonable outline of the OD region has become apparent
and therefore using the remaining white pixels, the center of the OD region has
been detected and an approximated boundary has been set. Depending on the
size of the OD in relation to the overall retinal image size and capturing device
157
8. OPTIC DISK AND MACULA LOCALISATION AND EXTRACTION
specifications, this boundary might vary and therefore has to be set accordingly.
In this case the given boundary has a radius of 70 pixels.
Using the centre location and plotting an approximate circle a mask has been
created to locate the OD region.
Implementing the mask on to the original image has provided a more specific
region of interest. Since consecutive application of a process would enhance the
accuracy and speed of detection, at this stage the thresholding has once again been
applied. The result for the overall process has been displayed in the next section.
8.3 Implementation
In this section, the proposed consecutive adaptive thresholding technique for de-
tection of the OD has been implemented and the results have been displayed in
Table 8.1. It should be noted that for clearer visibility of the results, the images
were zoomed in.
The Adaptive thresholding method has been performed twice on the desired
image in order to accurately detect its brightest regions or in other words OD.
Table 8.1a is an example of a possible desired retinal image which has been
pre-processed according to the procedure covered in Chapter 5 and has been used
in this section for OD localisation.
Table 8.1b represents the detected brightest regions of the original image which
have the pixel values greater than 200 gray levels and are considered as the upper
region of the image histogram. It also outlays the regions which are most likely to
be the OD.
The next two rows, Table 8.1c and 8.1d represent the regions which are having
the pixel values greater than the calculated mean value and the minimum value of
the upper region respectively.
Combining the findings would result in detection of the possible center of the
OD region and is shown in Table 8.1e.
Once the center has been localised, a boundary is set and plotted to the region
which is most likely to be the location of the OD. This has been displayed in
Table 8.1f.
158
8. OPTIC DISK AND MACULA LOCALISATION AND EXTRACTION
Table 8.1: Step by step results for OD detection, applying the proposed consecutiveadaptive thresholding method.
STAGE 1
(a) Original image
(b)Detected lighter region with pixel values greater
than 200 gray level
(c)
Detected lighter region using pixels with values
greater than the mean value from the upper region
of the histogram
(d)
Detected lighter region using the pixels with val-
ues greater than the minimum value from the up-
per region of the histogram
(e)Detected center of OD region using results from
part(d)
(f) Outline of the OD region
STAGE 2
(g)Cropped OD region using results from the first
cycle
159
8. OPTIC DISK AND MACULA LOCALISATION AND EXTRACTION
Interested Region Outcome
(h)Detected lighter region with pixel values greater
than 200 gray level outlined in blue
(i)
Detected OD using the pixels with values greater
than the minimum value from the upper region of
the histogram outlined in green
(j)
Detected OD using the pixels with values greater
than the mean value from the middle Region of
the histogram outlined in red
(k)Detected OD by illustrating results from sections
(h), (i) and (j).
The possible OD region has now been determined. In order to segment the OD
region in the original retinal image, a mask has been created and implemented and
the outcome is shown in Table 8.1g. At this point a second round of the adaptive
thresholding procedure has been applied.
Similar to the previous round, the Table 8.1h, 8.1i and 8.1j represent the de-
tected lighter regions with pixel values greater than 200 gray levels, minimum value
of the upper region and mean value of the middle region accordingly.
Combing the results and plotting boundaries around the detected regions, out-
lines the possible location of the OD. The final result is illustrated in Table 8.1k.
This automated process has proven to be successful in localising the exact
boundary of the brightest region of the retinal image, which is considered as the
OD. It has been implemented on more than twenty different images. The re-
sults obtained using this exact OD detection methodology appears to be of higher
precision in comparison to the other available procedures, with an average com-
putational time of 20-25 seconds.
160
8. OPTIC DISK AND MACULA LOCALISATION AND EXTRACTION
Table 8.2: OD localisation for different images.
Original
Image
Pixels > min-
imum value of
upper region
Pixels >
mean value of
middle region
OD
Image 1
Image 2
Image 3
Image 4
Image 5
Table 8.2 is the results of five different images from the studies database, illus-
trating the OD localisation using the proposed Consecutive Adaptive Thresholding
technique. The results show that the OD localisation has been successful for all
cases and the OD boundaries have been exactly detected.
161
8. OPTIC DISK AND MACULA LOCALISATION AND EXTRACTION
8.4 Optic Disk Extraction
Once the OD has been localized, the next step is to extract the necessary needed
information from the detected ROI.
During disease diagnosis and its progression, the ophthalmologists look at the
shape of the OD and its variation in diameter. Therefore, estimation of the OD
center is the first necessary step in extracting the information. This is then followed
by calculation of its area and later on determing the cup to disk ratio.
8.4.1 Center of the Optic Disk
Two approaches have been suggested by the author, in order to estimate the
location of the center of the OD.
The first method is to estimate the location of the center of the OD based
on the detected boundary of the localised OD. This is similar to the previously
suggested center calculation for Iris and Pupil in Section 6.4.1 of this chapter.
The second method is to determine the location of the center based on the
originating of the vessels within the OD. In this case the use of Template Matching
has been suggested.
1. Extract the OD from the retinal image.
2. Localize the vessels within the OD using the methodology suggested in sec-
tion 7.1 of this chapter.
3. Mask the vessels, so that the vessels have the pixel value of 1 and the re-
maining areas of the OD have the 0 pixel value.
4. Trace the location of the vessels. If necessary burst or shrink the vessels to
the desired thickness.
5. Calculate the point of intersection by implementing template matching con-
cept, determining whether the surrounding pixels of a middle value pixel in
a 3×3 window is zero or one. If the middle pixel in red is our desired pixel
and it is surrounded with at least three other non-zero pixels as shown in
162
8. OPTIC DISK AND MACULA LOCALISATION AND EXTRACTION
Figure 8.5, it can be said that intersection has occurred at the desired pixel.
Otherwise, it can be the middle or an end point in the vessel.
Figure 8.5: Some possible templates for determining vessels intersection
6. The point of intersection, represents the origination of the vessels and so the
center of the OD.
Example of the result obtained implementing the suggested methodologies to
detect the center of the OD can be viewed in Figure 8.6.
(a) Detected OD center
;(b) Zoomed in image
;
Figure 8.6: Center localisation of the OD, method 1 is represented as a blue (+)sign and method 2 as red (+) sign
Comparing the two methodologies, the results are approximately similar. In
majority of the cases, the first methodology is sufficient, unless otherwise the
location of the origination of the vasculature is also on importance in disease
detection and prognosis. The overall processing time is about 1-2 seconds and this
is due to simplicity of the process and reduction in the size of the ROI by confining
it to the OD region.
163
8. OPTIC DISK AND MACULA LOCALISATION AND EXTRACTION
8.4.2 Area of the Optic Disk
Area of the OD can be calculated following the same procedure which was used
previously in Section 6.4.2 for determining the area of the Iris and Pupil.
8.4.3 Cup to Disk Ratio
For diagnosis of diseases such as Glaucoma, ophthalmologists would consider the
area and variation in the shape of the OD, as well as the cup to disk ratio [46, 157,
158, 159]. In the previous sections, the area and the overall shape of the OD has
been detected and analysed. In this section, a suggestion has been made to detect
the Optic Cup (OC) so that it could be used to determine the cup to disk ratio.
Detection of the OC which outlines the borders of the Optic Nerve Head (ONH)
is quite difficult in comparison to the OD localisation as it may not clearly be visible
in the fundus image. On the coloured retinal images, it usually appears as a pink
colour or change in contour from rim to the cup [160].
Although, it may not be possible to accurately detect the OC in all the images
as it may not be visible, in this study it has been suggested to detect the OC using
the similar approach as the suggested consecutive adaptive thresholding which was
used for OD localisation. The overall procedure would be similar but performed
on contrast enhanced images. Since the contrast of the images has changed, the
automatically detected threshold value would also defer, resulting in detection of
the OC. Enhancing the contrast would help in distinguishing and detection of the
cup boundary. More details on how to enhance the contrast of retinal images is to
be covered in Chapter 5.7.
Figure 8.7: Detection of the OC (green) and OD (red)
164
8. OPTIC DISK AND MACULA LOCALISATION AND EXTRACTION
Figure 8.7 represents a sample result which have been obtained when imple-
menting the procedure to detect the OC and OD. As it can be seen, the OC has
been detected and the boundary is shown in green, while the detected OD has
been shown in red.
Once the cup has been localized, similar feature extraction procedures as those
for OD can be performed in order to determine the radius and area of the cup.
Using the obtained area, the ratio between the OD and OC can then be calculated
and used by ophthalmologists for determining the rate of progression of diseases.
The common approach is to visually examine the symmetrical and shape of the OD
and OC under the slit lap biomicroscopy. In the case of Glaucoma, based on the
study performed by Nicolela [160], the cup to disk ratio asymmetry of 0.2 or greater
between the fellow eyes of the patient can be suggestive of this disease.. Therefore,
with the aid of the suggested approach it is possible to help ophthalmologists with
their diagnosis.
8.5 Macula Localisation - Proposed Technique
In order to detect the Macula, its visual characteristics have to be defined. Based
on the definition mentioned in section 2.1.2, Macula is a darkly pigmented circular
region near the center of the fundus retinal images and its structures are responsible
for high acuity vision.
(a) (b) (c) (d)
Figure 8.8: Different positions of macula in retinal images, in images (a) and (d)macula is located in the center while in images (b) and (c) no macula is present.The macula has been manually defined and can be viewed in the images.
165
8. OPTIC DISK AND MACULA LOCALISATION AND EXTRACTION
From observing a number of fundus images, it can be seen that macula may or
may not be present in the image of interest, as shown in Figure 8.8. In Figures 8.8b
and 8.8c, no Macula is visible because of the angle of the image. It can also be
seen that depending on the location of the OD, macula may approximately be
localized as well.
Therefore in this study, prior to localisation of macula, it has been suggested
to initially locate the OD. This is then followed by defining whether Macula is
expected to be present or not. In cases where the Macula is not expected to be
present, further processing is not necessary. However, in cases where the Macula
is expected to be present, process proceeds and macula is localised using Neural
Network (NN) concept.
Figure 8.9 illustrates the flowchart of the overall procedure of the proposed
technique by the author for detection of Macula.
EyeImage
Acquisition
Image
Pre-Processing
Macula
Localisation
OD Coordinates
Is Macula Present?
Complement Image
Adaptive Thresholding
No Further Analysis
Interpretation
(Feature Extraction)Display
Center
Radius
AreaNoYes
Figure 8.9: Proposed steps for Macula localisation.
Neural Networks has been widely used in different areas. In ophthalmology, it
has mainly been used in detection of vessels in retinal images [161]. However in
this study, the concept of NN has been used to determine whether Macula is or
is not present in the given retinal image. If it is present, the Macula can then be
localised.
166
8. OPTIC DISK AND MACULA LOCALISATION AND EXTRACTION
To do so, the fundus image has been segmented in to blocks. 9 blocks in this
case. The number of blocks depends on the ratio and the size of OD in relation
to the size of the retinal image. Based on the observation, for the set of analysed
image in this case, 9 blocks have been sufficient and resulted in accurate macula
localisation. In other cases, in which the image or the capturing instrumentation
specification may vary, the number of blocks may also vary. Figure 8.10 illustrates
a sample of retinal image being separated into the desired number of blocks.
Figure 8.10: The retinal image has been deperated into blocks.
Using the simple feed-forward concept of the NN depicted in Figure 8.11, each
block is considered as the input. The inputs are then checked for the presence
of the OD with them. If OD is present, the output would be set as 1, otherwise
it would be set to 0. It should be noted that the weight for each input block is
the same since OD may be present in any of the blocks. Once the block in which
contains the OD is determined, the blocks which are most likely to contain the
Macula are then investigated further.
167
8. OPTIC DISK AND MACULA LOCALISATION AND EXTRACTION
Input Layer Hidden Layer Output Layer
Input 1
Input 2
Input 3
Input 4
Input 5
Input 6
Input 7
Input 8
Input 9
OD
No OD Ouput
Figure 8.11: Neural network model determining the OD block.
Based on observations, the OD is normally located in the centre, sides or diag-
onals of the images. Depending on the number of blocks and the location of the
OD, it is then possible to estimate the location of the Macula.
Moreover, on average Macula is approximately located 3 mm temporal to the
168
8. OPTIC DISK AND MACULA LOCALISATION AND EXTRACTION
OD [162]. Therefore, on images where OD is located on the right or left hand
side, Macula is most likely located in the center of the image. However, this is still
dependant on the magnification and the angle in which the image is taken from.
If the instrument or possible output of the retinal fundus image is unknown,
it is best to implement Macula detection process to all the blocks expect those
of the background and the OD. This would ensure that the macula is detected
irrespective of the possible estimated ROI.
OD has been previously detected in Section 8.1. To detect Macula, it is impor-
tant to determine in which of the created blocks the Macula is more likely to be
present. For example in Figure 8.10, since the OD is located on the left hand side
in block (4), the Macula is most likely be present in center of the image in Block
(5) or with a lower probability on the right hand side in block (6). Therefore, the
localisation process may only be applied to these two blocks.
There are also possible cases where the OD is not present. In such cases the
Macula localisation may proceed throughout all the blocks.
In other cases where the OD is present in the middle block (block 5), the Macula
may or may not be apparent in the image and therefore the Macula localisation
procedure has to be implemented to all the blocks. However, there is a possibility
that the Macula is covered by the OD and may not be visible.
Once the possible blocks for which the Macula is most likely to be presented in
has been defined, the processing steps similar to those previously used for OD lo-
calisation can be implemented. However, there is a slight alteration to the method-
ology.
Since the Macula is a dark circular region of the retina, the darkest pixels
have to be located instead of the brightest pixels which have been previously
selected in the case of the OD. Another option which was implemented in this
study would be to obtain the complementary image, in which the brightest pixels
correspond to the darkest pixels of the original image or vice versa. An example of
the complementary image can be viewed in Figure 8.12. Once the complementary
image is obtained, the localisation process would be exactly as it was for the case
of the OD.
169
8. OPTIC DISK AND MACULA LOCALISATION AND EXTRACTION
(a) Image (b) Complement
Figure 8.12: Complementary image. (a) Original Image, (b) Complement Image.
Advantages of using the proposed technique are that the error in localisation
of the Macula is reduced significantly. The error is reduced as the first step is to
deter ermine the presence of the macula. If this step is not included the macula
may be located wrongly. Moreover, since the desired ROI is reduced in size the
overall processing time has also reduced. Detection of both Macula and OD may
also be helpful in more accurate formation of fundus maps, which was discussed
in chapter 3, as these features can also be used as markers similar to the vessels
locations.
8.6 Implementation
The proposed methodology for Macula localisation has been performed and the
outcome can be viewed in Figure 8.13. The results are promising and the approx-
imate detection of the Macula has been a success.
(a) Image (b) Detected Macula
Figure 8.13: Localisation of Macula using the proposed technique.
170
8. OPTIC DISK AND MACULA LOCALISATION AND EXTRACTION
Table 8.3: Macula localisation for different images. For cases where the Maculacannot be seen the process is stopped, such as the case for Image 6.
Image Complement Macula Region Macula
Image 1
Image 2
Image 3
Image 4
Image 5
Image 6Macula not vis-
ible
Macula not vis-
ible
171
8. OPTIC DISK AND MACULA LOCALISATION AND EXTRACTION
Observing the results suggests that Macula detection using the proposed tech-
nique has been successful. This process has been implemented on over twenty
images from the used databases with an average processing time of about 3-5
seconds, some of the results have been illustrated in Table 8.3.
8.7 Macula Extraction
Macula is considered to be circular in shape, similar to the OD. Therefore, the
important information which may need to be extracted from the Macula is the
center and the radius. As a result, the approaches undertaken to estimate these
information are the same as what was previously suggested for OD extraction.
Since in this case, the origination of the blood vessels were not of interest and
so the chosen process for locating the center of macula was similar to the one
suggested in Section 6.4.1.
The approach for calculating the area of the macula was also similar to the
suggested method in Section 6.4.2.
The approximate processing time for detection of the center and area of the
macula was less than 1 second, which suggests a very fast processing time due to
simplicity of the suggested procedures.
8.8 Summary
Localisation and extraction of OD and Macula has significant impact in ophthal-
mology as some of the widely affecting diseases such as Glaucoma affects these
features. Therefore variation in shape of OD and Macula can be useful in an early
detection of these diseases.
This chapter looked into the possibility of extracting information from these
features via their accurate localisation. A new method of Consecutive Adaptive
Thresholding technique has been introduced for finding the brightest pixels in
the image in order to exactly outline the OD boundary with a high accuracy.
Possibility of detecting OC has also been suggested at this section, as knowing
the ration between the OD and OC is used to determine the possibility of the
172
8. OPTIC DISK AND MACULA LOCALISATION AND EXTRACTION
occurrence of a disease.
Similarly Macula has also been localised using the same technique. However,
since Macula is the darker region of the retina, the process was altered to some
extent. The Consecutive Adaptive Thresholding approach in this case was used to
detect the darker region of the retina on the complement image. Moreover, there
are times where Macula is not visible in the image as it is over shadowed by the
OD. Hence, prior to the implementation of the technique, Neural Network concept
was applied to determine whether Macula was present or not. If it was present,
then the procedure was performed.
To extract information from the localised OD and Macula, their center was
initially detected. This was then followed by radius and area of the two regions.
In the case of the OD, the ration of the OD to OC was also determined as the
determining factor of occurrence of diseases such as Glaucoma.
The proposed new approach was able to accurately locate the OD and Macula.
The exact boundary detection instead of circular assumption was performed in
order to enhance the accuracy of the extracted information further. The extracted
features were also calculated in order to help ophthalmologists in their diagnosis.
173
CHAPTER 9
CONCLUSIONS
Introduction Literature Review Thesis Outline Conclusion
Figure 9.1: Chapter Nine Outline
9.1 Overall Research Program
Ophthalmology has been a growing field in the recent years. With the aid of
the new medical instrumentations and Telemedical devices, ophthalmologists have
been able to diagnose, treat and monitor patients.
The most important stage for treatment of any condition is its early detection.
To aid the ophthalmologists in the diagnosis stage, this study concentrated on
some of the most widely affecting disease such as Cataract, Glaucoma, and ROP.
For each of their key descriptors and features; Iris and Pupil, OD and Macula and
retinal vessels; image processing techniques were suggested for their localisation
and examination. Furthermore, the study was designed such that it could be
used as part of a Telediagnostic tool, which could also be used in rural areas and
developing regions where the availability of resources and expertise are limited.
To achieve this objective, improvements and modifications for all stages of
174
9. CONCLUSIONS
image processing, including image acquisition, pre-processing, feature localisation
and extraction were suggested. The considerations for the proposed techniques
were the simplicity, robustness, fast processing time and high accuracy, with min-
imal user input. The processes were designed such that all the obtained results
could be stored onsite or transferred offsite to be used by ophthalmologists for
their prognosis.
9.2 Research Findings, Perceived Contributions
In this study, each chapter has concentrated on a specific stage of image processing
and some modifications were suggested for each stage. The proposed techniques
were fast, reliable, non-invasive and with a reasonable accuracy.
In order to examine and study a problem, data is required. Therefore, the
first step of image processing is image acquisition. For the purpose of this study,
open source data bases were chosen so that the compatibility of the procedures on
different input data could be monitored and examined. Several different databases
were considered including STARE, DRIVE, MESSIDOR, REVIEW, ROC, CMIF
and UPOL. To study the Iris and Pupil, the images from UPOL database were
chosen. For the cases were retinal fundus images were required, the STARE and
DRIVE data bases were chosen because they are the most widely used databases
by researchers in this field were chosen. All the consecutive steps in this study
were performed on over twenty different images from these databases.
Due to limitations, accessibility and cost of instrumentations in remote loca-
tions, majority of developing nations may only have access to minimal resources.
As a result, to create a wider view on the retina, the use of multiple markers and
images was suggested in order to create a fundus map using normal view angle
cameras. Ophthalmologists can then use this map to diagnose and treat diseases.
The suggested methodology for creating the fundus map used geometric charac-
teristics of the images and included overlapping regions with more markers. As a
result, significant amount of unwanted duplicate noise was removed.
The next a crucial step in image processing is image pre-processing. Different
stages including the colour separation, segmentation and masking of the ROI, noise
175
9. CONCLUSIONS
removal of the images using 2D FFT filters and sharpening them were considered
and examined. These steps were performed in order to prepare the images for the
consecutive stages, their main objective is to remove all the unwanted information
and as a result reduce the overall processing time.
Moreover, studying the results indicated that in some special cases further
image pre-processing stages may be required. Two of such circumstances were
studied as part of this research.
Observing the results indicated that there were times where the bright fringe
noises affected the detection precision, especially in the case of OD localisation.
As a result, circular and elliptical trimming was suggested to be implemented
prior to feature localisation, in the pre-processing stage. After this application,
the precision of results performed was greatly improved to 100% success rate in
comparison to the other previously suggested procedure in the literature.
In other cases, the accuracy of the results was affected due to the contrast
of the images. This specially became apparent when thresholding technique was
considered. In such cases, the contrasts of the images were enhanced using the
Intensity Adjusted, Histogram Equalization and Adaptive Histogram Equalization.
The next two main stages of image processing included the feature localisation
and extraction. As mentioned previously, the main key features considered in this
study for diagnosing Cataract, ROP and Glaucoma were the Iris and Pupil, retinal
vessels, OD and Macula respectively.
For detection of the Iris and Pupil boundaries, which is beneficial for Cataract
diagnosis and Biometrics application, an amalgamates procedure was suggested to
incorporate and combine the results from two or more different processes in order to
create a single outlay to mask and segment the ROI. By combining two different
techniques of Thresholding and Active Contouring, the suggested methodology
has improved the accuracy of the detection in about 2-5 seconds. Since the two
procedures occur simultaneously the processing time is not increased significantly
while the results are of higher precision. To quantize the chosen region approaches
were suggested to approximately determine the center of the Iris and Pupil and
then calculate their areas.
Using fundus images, the retinal vasculatures were examined and localised for
diagnosis of diseases such as ROP, where the complication affects the vascular for-
176
9. CONCLUSIONS
mation and shape. To do so, two different size kernels of several filters were applied
to the readily available 2D FFT filtered images from the pre-processing stage. The
applied filters included the Sobel, Canny, Laplacian, Prewitt, Circular Average,
Average, Median, Weiner and Gaussian filters. The observed results suggest that
larger size kernels revealed more information, but had more unwanted noise as
well. From the implemented sample filters, four revealed more details. The filters
were the Average filter, Median filter, Weiner filter and Gaussian filter. However,
there was still noise present in the outcomes; hence a combination of them was
studied. Three of the results showed very clear vasculature edges, including the
application of Average filter or Gaussian filter on the Weiner filtered image and
Wiener filter when it was applied to the Median filtered image. There were some
slight variation between the three best results but by observation, it could be said
that when Weiner filter was applied on a median filtered image, some more details
could be viewed in the output. To analyse the findings it was suggested to de-
termine the end-point of the vessels using template matching. The simplicity and
the reasonable processing time of about 12-15 seconds for the suggested vascular
localisation process makes it a suitable preliminary telemedicine tool for determin-
ing the high risked patients who might suffer from retinal vascular disorders such
as ROP.
Lastly, OD and Macula are used to diagnose and monitor the progression of
Glaucoma. To localise the boundary of OD, a new iterative thresholding method-
ology was suggested. On the contrary to the majority of the available OD approx-
imation localisation techniques, this method determines the exact OD location
and shape. The variations and changes to the OD shape were also examined by
obtaining its center, area and cup to disk ratio. The overall processing time for
OD localisation was about 20-25 seconds.
Macula was examined using a similar approach as to the OD. Firstly, the
retinal image is checked for the visibility of the Macula using the Neural Network
concept. If the Macula was visible, the thresholding approach was applied to the
complement of the image, localising the Macula in 3-5 seconds. Center and area
of the Macula were also calculated.
This study indicated that if need be all the main key features for critical wide
spread diseases may be localised and monitored in under a minute. The simplicity
177
9. CONCLUSIONS
and robustness of the chosen approach also ensured that it could be used as part
of a Telediagnositc tool.
9.3 Proposals for Future Research
Research is an ongoing field and with advancements in technology, current available
approaches may further improve. In this section, a few suggestions have been given
by the author as the possible future work.
• Larger databases
– The databases used in this study were limited to the online available
open source links. However, by creating a larger database not only
the accuracy of detection for a specific disease increases but also wider
range of ailments could be detected using disease maching.
• Considering other diseases
– There are times when irreversible damage may be caused when critical
information is missed by the medical practitioner. This may be due to
the limited expertise of the ophthalmologists, rareness of a disease or
patient suffering from several medical conditions. In such cases, looking
into a larger database, covering many other diseases can be of great
assistance. To do so, many other studies on different diseases should
be performed so that the overlapping information and features of the
ailments could be defined and a larger database formed. Using this
database, then the medical practitioner could pinpoint and determine
what the main cause of the condition.
– One of the diseases which affect a wide population is Primary Angle
Closure Glaucoma (PACG). PACG causes development of angle closure.
The narrow angle is treated using the laser periphery iridotomy, if de-
tected early. Currently the Ultrasound Bio-Microscopy (UBM) is used
to detect the narrow angles, but since the procedure involves immersion
178
9. CONCLUSIONS
of the eye into fluid, it is time consuming, inadequate and inconvenient
as a routine test. To improve the testing procedure, a study can be per-
formed incorporating results from the Optical Coherence Tomography
(OCT) and the Ultrasound Bio-Microscopy (UBM). Since both OCT
and UBM are capable of obtaining cross sectional images of the Ante-
rior Chamber (AC), they may provide a better potential information in
detection of patients who might be at risk of angle closure.
• Considering other features
– In the image acquisition stage of the study, the effect of light and its
refraction was considered when passing through the eye and creating the
retinal images. The results suggested that there is a significant different
between the incident and refractive light rays which is usually ignored
and is not considered when analysing results. In order to consider this
effect, further calculation in consecutive steps of processing is needed
for determining the exact locations of the key features of the eye.
– Localization of other features and conditions can further help in cre-
ating a broader database. An example could be including results from
detection of microaneurysm [163].
– As mentioned previously, main feature for detection of the occurrence
of the complication during cataract surgery is variation in colour of
the eye. Using the colour index and its variation can therefore help in
detection of such complications.
– For the case of ROP, since the outgrowths of vessels are of great impor-
tance, the use of fractal approach can help in estimating the angiogene-
sis growth. This information can be included and used by the surgeons
to oversee the progression of the disease.
179
9. CONCLUSIONS
• Combining other results
– Patients Records
∗ The obtained results may also benefit if other information from the
patient such as ocular pressure could be available. For example if
variation of ocular pressure could be constantly monitored during
the cataract surgery, any changes in pressure may assist the pro-
posed monitoring system and alert the surgeons of the possibility
of complication.
– Results from other devices
∗ In this study, image processing methodologies were of main inter-
est. Incorporating the information and results from other medical
devices such as the OCT [164], fluorescence angiograms, use of in-
frared lighting with the results obtained from this study can aid
the ophthalmologists to make a more valid and reliable decision in
their disease prognosis.
∗ Real time feedback from the OCT can also aid the cataract surgery
significantly as the thickness of the posterior capsule can be con-
tinuously measured intra-operatively. Any changes in the thickness
can then alert the surgeon. This may also help in creating the 3D
view of the eye during the surgery as the location of the device, the
depth of the eye and all its features can be calculated and defined.
∗ Including results obtained from Confocal Scanning Laser Tomogra-
phy (CSLT) which is widely used for three dimensional scanning of
the ONH would provide a better insight into the extent of progres-
sion of Glaucoma. However, further statistical examination of the
progression of the structural glaucomatous damage as well as im-
provements on the repeatability of the images obtained using this
technique is required.
To do so, a Statistic Image Mapping (SIM) can be performed which
may benefit the field of neuro-imaging. The active changes of the
ONH can be visualised by applying the pixel by pixel analysis of
180
9. CONCLUSIONS
the topographic height over time. The flagged change map and
the intensity variation can be used to determine active changes
of the ONH and determine the progression of the disease. The
repeatability of the images can be tested by comparing the findings
with the results obtained from the Topographic Change Analysis
(TCA) system.
• Improvements of devices
– Hardware improvements - For instance in the cataract surgery case,
placing sensors on the head of the phacoemulsification handheld device,
in order to measure the input and output flow can help in constant
monitoring of the intraocular pressure and so automatically stoping the
surgery if any irregularity is seen.
– Improvements on portable handheld capturing devices - With increase
in technology and its availability in remote locations, image processing
can further enhance. Capturing high quality retinal images using mo-
bile phones are the next step in disease classification. Despite several
studies being recently conducted in this field, it may still acquire im-
plementation of several new filtering systems and image enhancement
mechanisms.
– Improvements of OCT -
∗ Create a real time, high speed anterior segment OCT system which
can quantitatively analyse the angle parameters. The designed
OCT should use the 1.3µm light source instead of the 0.8µm light
source which would provide better visualisation of the features and
enhance the speed of processing significantly faster than the current
available segment OCT systems. This is due to the lower scattering
of light at this wavelength as well as about 90% reduction of light
incident reaching the retina as it is absorbed by the water in the
ocular media. This system may be applied to analyse the angle
parameters, which can then be used for narrow angle detection and
diagnosing diseases such as PACG.
181
9. CONCLUSIONS
∗ Incorporation of the ultra-board spectral bandwidth light sources
in order to reduce the cost and enhance the axial resolution of
OCT production. The OCT technology may further benefit from
combining the outputs of other available technologies such as the
[148] N. OTSU, “A threshold selection method from gray-level histograms,” IEEE
Trans. Syst. Man Cybern, vol. 9, pp. 62–66, 1979.
[149] H. Ng, “Automatic thresholding for defect detection.” Pattern Recognition
Lett., vol. 27, p. 16441649, 2006.
[150] P. Richard, “Iris recognition: an engineering biometric,” Proceeding of the
IEEE, pp. 1354–1356, 1997.
[151] M. Brewer, “A modelling approach for bandwidth selection in kernel density
estimation,” COMPSTAT, pp. 203–208, 1998.
[152] Z.Zhang, F.S.Yin, J.Liu, W.K.Wong, N.M.Tan, B.H.Lee, J.Cheng, and
T.Y.Wong, “Origa-light : An online retinal fundus image database for glau-
coma analysis and research.” 32 nd Annual International Conference of the
IEEE EMBS, pp. 3065–3068, 2010.
[153] I. Kasa, “A curve fitting procedure and its error analysis.” IEEE Trans. Inst.
Meas., vol. 25, pp. 8–14, 1976.
[154] D. Chang and W. Wu, “Image contrast enhancement based on a histogram
transformation of local standard deviation,” IEEE Transactions on Medical
Imaging, vol. 17, no. 4, pp. 518–531, 1998.
[155] M. Aarthy and et al, “International journal of computer science and mobile
applications,” IEEE Transactions on Medical Imaging, vol. 2, no. 3, pp.
25–34, 2014.
[156] Zuiderveld and Karel, Contrast Limited Adaptive Histograph Equalization.
Graphic Gems IV. San Diego: Academic Press Professional, 1994.
[157] D. Grigera, “Clinical estimation of opric disc size,” Glaucoma Now, pp. 10–
11, 2011.
[158] S. Zeki, J. Dudgeon, and G. Dutton, “Reaooraisal of ratio of disc to mac-
ula/disc diameter in optic nerve hypoplasia,” Br J Ophthalmol, vol. 75, pp.
538–541, 1991.
198
REFERENCES
[159] J. Jonas, G. Gusek, and G. Noumann, “Optic disc, cup and neuroretinal rim
size, configuration and correlations in normal eyes,” Investigative Ophthal-
mology and Visual Science, vol. 29, no. 7, 1988.
[160] M. Nicolela, “Optic nerve: Clinical examination,” in Pearls of Glaucoma
Management, J. Giaconi, Ed. Berlin Heidelberg: Springer-Verlag, 2010.
[161] S. Manoj, Muralidharan, and P. M. Sandeep, “Neural network based classi-
fier for retinal blood vessel segmentation,” International Journal of Recent
Trends in Electrical and Electronics Engineering, vol. 3, pp. 44–53, 2013.
[162] K. Rohrschneider, “Determination of the location of the fovea on the fun-
dus,” Investigative Ophthalmology and Visual Science, vol. 45, no. 9, pp.
3257–3258, 2004.
[163] G. Rajput, P. Patil, and R. Chavan, Automatic detection of microaneurysms
from fundus images using morphological operations. Springer, 2013.
[164] D. Koozekanani, K. Boyer, and C. Roberts, “Retinal thickness measurements
from optical coherence tomography using markov boundary model,” IEEE
Transactions on Medical Imaging, vol. 20, no. 9, 2001.
199
APPENDIX A
Table ?? was used to plot the Figure 4.9. It was calculated using the Equation 4.12
indicating the relation between the index value of the angle of incidence and re-
fraction when light passes through two different materials.
n1sin(θ1) = n2sin(θ2) (1)
Table 1: Angle of light as it enters the eye (Incident Ray), passes through differentinterfaces within the eye and reaches the back of the eye (Refractive Ray).
Incident
Ray
(Degrees)
First
Interface
(Radians)
Second
Interface
(Radians)
Third
Interface
(Radians)
Fourth
Interface
(Radians)
Refractive
Ray
(Degrees)
0 0.00 0.00 0.00 0.00 0.00
5 0.07 0.06 0.06 0.07 3.74
10 0.13 0.13 0.12 0.13 7.46
15 0.19 0.19 0.18 0.19 11.16
20 0.26 0.25 0.24 0.26 14.82
25 0.32 0.30 0.30 0.32 18.43
30 0.37 0.36 0.36 0.37 21.96
35 0.43 0.41 0.41 0.43 25.40
40 0.48 0.46 0.46 0.48 28.74
45 0.53 0.51 0.50 0.53 31.93
50 0.57 0.55 0.54 0.57 34.96
200
. APPENDIX A
Incident
Ray
(Degrees)
First
Interface
(Radians)
Second
Interface
(Radians)
Third
Interface
(Radians)
Fourth
Interface
(Radians)
Refractive
Ray
(Degrees)
55 0.61 0.59 0.58 0.61 37.78
60 0.65 0.62 0.62 0.65 40.37
65 0.68 0.65 0.64 0.68 42.68
70 0.70 0.68 0.67 0.70 44.66
75 0.72 0.70 0.69 0.72 46.26
80 0.74 0.71 0.70 0.74 47.44
85 0.75 0.72 0.71 0.75 48.17
90 0.75 0.72 0.71 0.75 48.41
95 0.75 0.72 0.71 0.75 131.83
100 0.74 0.71 0.70 0.74 132.56
105 0.72 0.70 0.69 0.72 133.74
110 0.70 0.68 0.67 0.70 135.34
115 0.68 0.65 0.64 0.68 137.32
120 0.65 0.62 0.62 0.65 139.63
125 0.61 0.59 0.58 0.61 142.22
130 0.57 0.55 0.54 0.57 145.04
135 0.53 0.51 0.50 0.53 148.07
140 0.48 0.46 0.46 0.48 151.26
145 0.43 0.41 0.41 0.43 154.60
150 0.37 0.36 0.36 0.37 158.04
155 0.32 0.30 0.30 0.32 161.57
160 0.26 0.25 0.24 0.26 165.18
165 0.19 0.19 0.18 0.19 168.84
170 0.13 0.13 0.12 0.13 172.54
175 0.07 0.06 0.06 0.07 176.26
180 0.00 0.00 0.00 0.00 180.00
201
APPENDIX B
Following are the results obtained from conversion of the coloured images to their
corresponding gray scaled and indexed images.
Table 2: Gray Scaled and colour component separation of coloured images
Image Original Gray Scaled Red Band Green Band Blue Band
1
2
3
4
202
. APPENDIX B
Image Original Gray Scaled Red Band Green Band Blue Band
5
6
7
8
9
10
11
12
203
. APPENDIX B
Image Original Gray Scaled Red Band Green Band Blue Band
13
14
15
16
17
18
19
20
204
APPENDIX C
Following are the results obtained for the suggested approach in creating individual
masks for different images using the thresholding technique.
Table 3: Masks created for different images using Thresholding technique
ImageCreated
Mask
Masked Im-
ageImage
Created
Mask
Masked Im-
age
1 2
3 4
5 6
7 8
205
. APPENDIX C
ImageCreated
Mask
Masked Im-
ageImage
Created
Mask
Masked Im-
age
9 10
11 12
13 14
15 16
17 18
19 20
206
APPENDIX D
Following are the results obtained by implementing the 2D FFT filter.
Table 4: 2D FFT filtered images.
Image Gray Scaled Magnitude and Phase Plot Filtered
1
2
3
4
207
. APPENDIX D
Image Gray Scaled Magnitude and Phase Plot Filtered
5
6
7
8
9
10
11
12
208
. APPENDIX D
Image Gray Scaled Magnitude and Phase Plot Filtered
13
14
15
16
17
18
19
20
209
APPENDIX E
Following are the results obtained for sharpening the retinal images.
Table 5: Sharpening the retinal images using 2D FFT filtered images.
Image Gray Scaled FilteredSharpened-
10×5 kernel
Sharpened-
3×2 kernel
1
2
3
4
210
. APPENDIX E
Image Gray Scaled FilteredSharpened-
10×5 kernel
Sharpened-
3×2 kernel
5
6
7
8
9
10
11
211
. APPENDIX E
Image Gray Scaled FilteredSharpened-
10×5 kernel
Sharpened-
3×2 kernel
12
13
14
15
16
17
18
212
. APPENDIX E
Image Gray Scaled FilteredSharpened-
10×5 kernel
Sharpened-
3×2 kernel
19
20
213
PLEASE NOTE
The following materials cannot be reproduced online and have been extracted: Ektesabi, A & Kapoor, A 2011, 'Exact pupil and iris boundary detection,' Control, Instrumentation and Automation (ICCIA), 2011 2nd International Conference on, 1217-1221. DOI: 10.1109/ICCIAutom.2011.6356835 Ektesabi, A & Kapoor, A 2012, 'Complication prevention of posterior capsular rupture using image processing techniques,' Proceedings of the World Congress on Engineering, 603-607. www.iaeng.org/publication/WCE2012/WCE2012_pp603-607.pdf Ektesabi, A & Kapoor, A 2014, 'Removal of Circular Edge Noise of Retinal Fundus Images,' Proceedings of the International Conference on Image Processing, Computer Vision, and Pattern Recognition (IPCV). http://world-comp.org/preproc2014/IPC3384.pdf