Application of image processing techniques to tissue texture analysis and image compression

Application of image processing Application of image processing techniques to tissue texture techniques to tissue texture

analysis and image compressionanalysis and image compressionAdvisor : Dr. Albert Chi-Shing CHUNG

Presented by Group ACH1Presented by Group ACH1(LAW Wai Kong and LAI Tsz Chung)(LAW Wai Kong and LAI Tsz Chung)

Computer Science Final Year Project 2004

OverviewOverview

• Introduction– Motivation– Objectives

• Results– Classification algorithms:

• Feature extraction & Classifier selection

– Software implementation:

• Conclusion • Future Extension• Question and Answer Session

IntroductionIntroduction - Motivation- Motivation

Diagnosis of cirrhosis:

1) Manual diagnosis of ultrasonic liver image

2) Histological analysis •Invasive

•Inaccurate •Results dependent on experience of sonographers

Both are time consuming

How about computer aided diagnosis system?

In what extent this system assist doctor?

- Objectives - Objectives 1. Designated user interface with support of ultrasonic image

compression•No pre-image processing is needed

•Reduce storage space

Facilitate the diagnosis process

2. Multi-severity level classification

•Cirrhosis treatment require severity information.

3. Machine independence

•Compatible with different ultrasound scanning machine

Challenge !! How to classify patients?

2 steps

Step 1: Feature ExtractionStep 1: Feature Extraction

Firstly, extract useful features from image.

We have examined several feature extraction approaches for performance comparison

The most accurate approach will be implemented in our system

1. Direct comparison of wavelet coefficient(Haar, Symlets, Daubechies)

2. Histogram of wavelet coefficient (Haar, Symlets, Daubechies)

3. Statistic with “Difference on Gaussians” filter

4. Direct comparison between multi-scale co-occurrence matrix

}),(,),(,,

:)],(),,{[(1

),(

jnmgilkgdynldxmk

ROInmlkyCardinalitN

jiC

5. Statistic with multi-scale approach and co-occurrence matrix

Step 1: Feature ExtractionStep 1: Feature Extraction

The six features:

1) The mean gray level

ROIji

jiIN

mean),(

),(1

- Inversely proportion to cirrhosis severity. - Affected by the area of normal tumor

2) The first percentile of the gray level distribution P

P

ii

P

jj hh

0

1

0 10

1

First order statistic

- Inversely proportion to cirrhosis severity. - Affected by the present of normal tumor

Co-occurrence matrix statistic

3) Entropy:

Gji

jiCjiCEntropy,

)],(log[).,(

4) Contrast:

Gji

jiCjiContrast,

2 ),(.)(

5) Angular Second Moment:

Gji

jiCASM,

2),(

6) Correlation

yx

Gjiyx

SS

mmjiijC

nCorrelatio

,

),(

iy

jy

jx

ix

ijy

jix

mjiCjS

mjiCiS

jiCijm

jiCim

222

222

),(

),(

),(

),(

Features Relation of feature and cirrhosis status

Physical meaning

Entropy Inversely proportion Randomness of intensity changes

Contrast Inversely proportion Edge detection

ASM Inversely proportion Homogeneity of image

Correlation Proportion Similarity among pixel pairs

6) Morphological based method

• Segment out tumor structure from liver• Count the number and circumference of tumor

• Input features: normalized to range between [0,1]• Category: normalized to range between [0,1]• Classification: by setting thresholds base on # category. • 1st layer: 5 hyperbolic tangent sigmoid transfer units• 2nd layer: 1 linear transfer unit• Train function: Levenberg-Marquardt back-propagation• Performance: MSE• Stopping threshold: 0.01• Maximum training cycle = 200

Step 2: ClassifierStep 2: Classifier

• Basic requirements: – Continuous learning

– Multi class classification (severity category)

– Robust

– Database can update per patient (one pattern). Secondly, classify patients based on extracted features

3 classifiers were examined

1) k-Nearest Neighbor Classifier

• Use the category of k-nearest neighbor in database to classify a new entry.

• The features are normalized by standard score.

• Distance-weighted.

• Choice of distance: SSD / KLD

• Physically, KLD measures relative entropy between PDF

2) Feed-forward Neural Network

• A direct continuation of the work on Bayes classifiers, which relies on Parzen windows classifiers.

Setting:

3) Probabilistic Neural Network

• It learns to approximate the PDF of the training examples.

)(1

/)1(

1

/)1(22

j

kjk

ki

N

i

wxN

i

wx eeoutput

• The input features are normalized by standard score.

• Commonly used in image feature classification

)(

)(log)()(

2

1121 kp

kpkpppKLD

Evaluation of algorithmsEvaluation of algorithms

Method of evaluating hypothesis: 10-fold cross validation (in MatLab)

Problem: Images of the same patient have similar features!

Solution: Use patient ID to partition the data set.

Problem: uneven class distribution in folds!

Solution: Partition the patients based on their category, ensure class distribution is similar to original data set.

The features:

•Theoretically, morphology is a descriptive feature, but, practically, fine tuning of parameters is needed.•Segmentation parameter (sigma of Gaussian filter, initial marker intensity) too sensitive to suit all testing cases•Number of tumors was unreasonably fluctuated. (tumors count ranged from 15 to 90)

Comparison of best results among all features sets with different classifier:

Features Set Classifier Accuracy

Type Setting Type Parameters 2 Class Classification

3 Class Classification

Plain wavelet coefficient

3 Level

Haar

KNN K=5 301/732 41.1202%

234/732 31.9672%

Histogram of wavelet coefficient

2 Level

Haar

kNN KLD, k=12 548/772 75.9003%

431/772 59.6953%

Statistic with “Difference on Gaussians” filter

Filtering along X-direction

kNN K=19 531/772 (72.541%)

434/772 (59.2896%)

PNN 447/772 (72.541%)

396/772 (54.0984%)

FFNN 497/772 (67.8962%)

442/772 (60.3825%)

Features Set Classifier Accuracy

Type Setting Type Parameters 2 Class Classification

3 Class Classification

multi-scale co-occurrence matrix

3 Resolution level

kNN KLD, k=3 312/515 60.5825%

219/515 45.5242%

SSD, k=2 284/515 55.1456%

211/515 40.9709%

Statistic of multi-resolution and co-occurrence matrix

2

Resolution Level

kNN SSD, k=19 614/732 83.8798%

511/732 69.8087%

PNN 607/732 82.9235%

497/732 67.8962%

FFNN 619/732 84.5628%

508/732 69.3989%

The data set is captured by Dr. Simon Yu, consultant and adjunct associate professor from Department of Diagnostic Radiology and Organ Imaging, Prince of Wales Hospital

Evaluation of algorithmsEvaluation of algorithmsThe classifiers:

Accuracy:>>> all of them have similar results. >>> Depends on features.

Running time (including partition for 732 testing cases):

Classifier 2 classes 3 classes

kNN 2s 2s

FFNN 67s 80s

PNN 7s 7s

Pros and Cons

k-NN

FastEasy to implement

Sensitive to class distribution of data set.Size of database is large and linearly increasing.

FFNNSize of database is a small constant.Robust

Training is slow. (> 40 times of k-NN) Should update per epoch to prevent noise.

PNN

Fast

Highly sensitive to class distribution of data set. Size of database increases linearly.

k-NN

ConclusionConclusion

• Developed a designated classification system that can contribute to medical aspect

• Examined different machine independent classification algorithms for multi-severity classification

• Proposed utilization of multi-resolution statistic with co-occurrence matrix for cirrhosis detection

• Realized machine learning and image processing techniques in a real life situation

• Explored the knowledge about cirrhosis and liver

Future ExtensionFuture Extension

• Clustering of features

• Fine tuning the parameters of morphological approach

• Histological findings of cases will be able to improve our system

Question and Answer SessionQuestion and Answer Session