Telehealth and Computer-aided Diagnosis

Telehealth and Computer-aided Diagnosis

By Juan ShanApril 2013

TelehealthTelehealth is the delivery of

health-related services and information via telecommunications technologies.

Telehealth could be as simple as two health professionals discussing a case over the telephone or as complicated as doing robotic surgery between facilities at different ends of the globe.

Telehealth and Computer-aided DiagnosisTelehealth is an interdiscipline that involes

medical image processing, networking, cybersecurity, machine learning, database management, and data mining.

Computer-aided diagnosis (CAD) systems use computer techniques to assist doctors to analyze data and make diagnosis.

Computer-aided diagnosis (CAD) systems could be able to serve independently as one terminal in the telehealth network in the future.

CAD for Breast Cancer• Breast cancer is the #1

leading cause of cancer death for women at ages 20 to 59 [1,2].

• 226,870 newly diagnosed cases and 39,510 deaths in the United States in 2012 [2].

• The earlier the cancers are detected, the better the patients are cured [1].

Fig. 1. Incidence rate of cancers for females

MammographyPreviously, the most effective modality for detecting breast

cancer is mammography. Limitations of mammography:

1. The radiation might be harmful for the patients and radiologists.

2. High false positive rate: 65%–85%.

3. Hardly detect breast cancer for dense breasts.

Fig. 2. Mammography images

Ultrasound ImagingBreast ultrasound (BUS) imaging is superior to the

mammography:

1. Having no radiation, safer than mammography for patients and radiologists in daily clinical practice [14]

.

2. Cost effective and portable.

3. More sensitive than mammography for dense breasts ( i.e., suitable for women younger than 35 years old[12]).

Computer-aided Diagnosis (CAD)Advantages of a CAD system:Calculate computational

features and statistical features, which cannot be obtained visually or intuitively by humans (doctors).

Minimize the operator-dependent nature of ultrasound imaging [4] and make the diagnosis process reproducible.

Computer-aided Diagnosis (CAD)1. Preprocessing: enhance the

image contrast and reduce speckle noise.

2. Segmentation: separate the lesion from the surrounding tissues.

3. Feature extraction: extract critical features to distinguish benign and malignant lesions.

4. Classification: use machine learning techniques to classify the lesion into benign or malignant types.

Computer-aided Diagnosis (CAD)Segmentation is the most

important step. Since many crucial features to

distinguish benign and malignant lesions are based on the shape and boundary of the lesion.

Example of benign and malignant lesions

Malignant

Benign

Lesion segmentation is important and… difficult as well!Due to the nature of ultrasound imaging, the breast

ultrasound (BUS) images are degraded by speckle noise, low contrast, blurred edges and shadow effect.

Automatic segmentation is a challenging task.

An automatic segmentation method

J. Shan, H. D. Cheng and Y. X. Wang, “Effective and Automatic Breast Ultrasound Image Segmentation Using L-Means Clustering”, Medical Physics. Vol. 39, Issue 9, pp. 5669-5682,2012 Sep

The proposed methodROI generationSpeckle reductionContrast

enhancementClustering

ROI Generation

(a) Original image (b) Binarized Image

Binarize the image into background and foreground, by an adaptively selected threshold

Delete the boundary-connected regions and noise regions.

(c) Image after binarization

(d) Image after region deletion

ROI Generation

Rank the regions.

The one with the highest score is considered as the lesion region. (e) Image of the winning

region

0

, 1,...,( , ) var( )n

n n

AS n kdis C C C

ROI Generation

ROI generation result

(a) Original BUS image (b) Binary image (c) Winning region

(d) ROI generation

The proposed method cont.

ROI generationSpeckle reduction [5]

Contrast enhancementNeutrosophic l-means

Speckle reductionAn effective and fast algorithm is used: speckle reducing anisotropic diffusion (SRAD) [5]

The proposed method cont.

ROI generationSpeckle reduction [5]

Contrast enhancement

Neutrosophic l-means

Fourier TransformSpatial domain Frequency domain

1-dimentional

From 1-D signal to 2-D image 2-D Log-Gabor filters defined in polar coordinates:

where k is related to the bandwidth of the filter and w0 is the center frequency of the filter. q0 is the orientation. sq defines the spread of the Gaussian orientation function.

6 orientations (0°, 30°, 60°, 90°, 120°, 150°) are chosen to cover the whole spectrum.

In each orientation, local phase feature LPA is calculated.

2 20 0

2 20

(log( / )) ( )( , ) exp ( )2(log( / )) 2

Gq

w w q qw q

k w s

PMO ImageWhich orientation should be

used?The LPA in the direction of the

edge can better characterize the structure than the LPAs in other directions.

Phase in max-energy orientation = PMO

2 2

1 1

( ( )) ( ( ))n n

s s

Eng e s o s

q

How to process in frequency domainto get PMO image

2 20 0

2 20

(log( / )) ( )( , ) exp ( )

2(log( / )) 2G

q

w w q qw q

k w s

1

1 '( )n

s

LPA phase sn

q

1( , ) || angle( '( , )) || || tan (imag( , ) / real( , )) ||phase x s I x s x s x s

' [ *cos(2* )] / 2phase π π phase

0 , 30 , 60 , 90 , 120 , 150( , ) ( , ), arg max ( , )PMO i j LPA i j Eng i j

q

q

2 2

1 1

( ( )) ( ( ))n n

s s

Eng e s o s

q

21 4( 0.5) 0 0.5 1 0.5< 1

PMO PMOPMO

PMO

2

2

2 0 0.51 2(1 ) 0.5< 1 PMO PMO

PMOPMO PMO

PMO image example

(a) ROI

(b) De-speckled image

(c) PMO image

(d) Enhanced PMO image

(a) (b) (c) (d)

The proposed methodROI generationSpeckle reduction [5]

Contrast enhancement

Clustering

ClusteringClustering is the partitioning of a data set into

subsets so that the data in each cluster have some common attributes.

Basic clustering:http

://home.deib.polimi.it/matteucc/Clustering/tutorial_html/AppletKM.html

The proposed methodROI generationSpeckle reduction [5]

Contrast enhancement

Clustering

Evaluation

Breast ultrasound databaseThe database is composed of 120 BUS images. 58

cases are benign, 62 cases are malignant. Every lesion is manually outlined by an experienced

radiologist. The manual delineations are served as the standard to evaluate the segmentation method.

Comparison is necessaryCompare the new method with:

A segmentation method using active contour model [21]

A segmentation method using level-set model [28]

A segmentation method using watershed model [70]

Result of segmentation methods

(a) The original image. (b) Manual delineation by radiologist. (c) Output of the method in [21]. (d) Output of the method in [28]. (e) Output of the method in [70]. (f) Output of the proposed method.

Area error metrics

Accuracy comparison

Future directionContinue the research on breast cancer

ultrasound, to find more reliable segmentation methods, to extract the features of tumors, and to train classifiers that can automatically classify tumors into benign/malignant.

Dental X-rays dataset, to detect tooth root and dental diseases.

Other type of medical images and explore the possible application of computer-aided diagnosis.

Thanks! Questions?

References1. Cheng, H.D., Shan, J., Ju, W., Guo, Y., and Zhang, L. Automated breast cancer detection and classification using ultrasound images: A survey. Pattern Recognition 43, 1 (2010), 299-317.2.Jemal, A., Siegel, R., Xu, J., and Ward, E. Cancer statistics 2010. CA Cancer J. for Clininicians 60, (2010), 227-300.3.Sahiner, B., Chan, H.-P., Roubidoux, M.A., Hadjiiski, L.M., Helvie, M.A., Paramagul, C., Bailey, J., Nees, A.V., and Blane, C. Malignant and benign breast masses on 3D US volumetric images: Effect of computer-aided diagnosis on radiologist accuracy. Radiology 242, 3 (2007), 716-724.4. Hwang, K.-H., H., Lee, J.G., Kim, J.H., Lee, H.-J. Om, K.-S., Yoon, M., and Choe, W. Computer aided diagnosis (CAD) of breast mass on ultrasonography and scintimammography. In Proceedings of 7th International Workshop on Enterprise Networking and Computing in Healthcare Industry, 2005, 187-189.5. Yu, Y. and Acton, S.T. Speckle reducing anisotropic diffusion. IEEE Trans. on Image Processing 11, 11 (2002), 1260-1270.6.Shankar, P.M., Piccoli, C.W., Reid, J.M., Forsberg, F., and Goldberg, B.B. Application of the compound probability density function for characterization of breast masses in ultrasound B scans. Physics in Medicine & Biology 50, 10 (2005), 2241-2248.7.Taylor, K.J.W., Merritt, C., Piccoli, C., Schmidt, R., Rouse, G., Fornage, B., Rubin, E., Georgian-Smith, D., Winsberg, F., Goldberg, B., and Mendelson, E. Ultrasound as a complement to mammography and breast examination to characterize breast masses. Ultrasound in Medicine & Biology 28, 1 (2002), 19-26.8.Zhi, H., Ou, B., Luo, B.-M., Feng, X., Wen, Y.-L., and Yang, H.-Y. Comparison of ultrasound elastography, mammography, and sonography in the diagnosis of solid breast lesions. J. Ultrasound in Medicine 26, 6 (2007), 807-815.9.Chang, R.-F., Wu, W.-J., Moon, W.K., and Chen, D.-R. Improvement in breast tumor discrimination by support vector machines and speckle-emphasis texture analysis. Ultrasound in Medicine & Biology 29, 5 (2003), 679-686.21. Madabhushi, A. and Metaxas, D.N. Combining low-, high-level and empirical domain knowledge for automated segmentation of ultrasonic breast lesions. IEEE Trans. on Medical Imaging 22, 2 (2003), 155-169. 28. Liu, B., Cheng, H.D., Huang, J., Tian, J., Liu, J., and Tang, X., Automated segmentation of ultrasonic breast lesions using statistical texture classification and active contour based on probability distance. Ultrasound in Medicine & Biology 35, 8 (2009), 1309-1324. 70. Zhang, M. Novel Approaches to Image Segmentation Based on Neutrosophic Logic. Doctoral Dissertation, Utah State University, 2010.