1. Automatic Classification for Pathological Prostate Images Based on Fractal Analysis Source: IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 28, NO. 7, JULY 2009 Author: Po-Whei Huang and Cheng-Hsiung Lee Report Date: 2011/12/16 Reporter: Hsin-Tse Lee

2. Outline Introduction Related Work Fractal Geometry Analysis Pathological Prostate Images Feature Extraction Extracting Feature by Differential Box-counting(DBC) Extracting Feature by Entropy-Based Fractal Dimension Estimation Combination of Two Fractal Dimension Texture Features Example 2

3. Outilne Classification And Feature Selection Classification Methods Estimation for Accuracy of Classification Feature Selection Experimental Results And Analysis Image acquisition Feature Sets Used for Comparison Performance of FD-Based Feature set Comparison of CCR Using Classifiers Without Feature Selection Comparison of CCE Using Classifiers With Feature Selection Discussion And Conclusion 3

4. Introduction(1/2) By viewing the microscopic images of biopsy specimens, pathologists can determine the histological grades. 4 Gleason grading diagram. Gleason grading is based upon the degree of loss of the normal glandular tissue architecture.

5. Introduction(2/2) 5 We can also see that the texture of prostate tissue plays an important role in Gleason grading for prostate cancer. (a)Gleason grade 2. (b)Gleason grade 3. (c)Gleason grade 4. (d)Gleason grade 5.

6. Fractal Geometry(1/2) 6

7. Fractal Geometry(2/2) 7

8. Analysis Pathological Prostate Images The first approach focused on the identification of the normal and abnormal tissue composition. Six texture features and two structural features were extracted from the image captured in each channel. The second approach focused on automatic Gleason grading for prostatic carcinoma. Extracted statistical and structural features from the spatial distribution of epithelial nuclei over the image area. 8

9. Extracting Feature by Differential Box- cunting(DBC) DBC is most commonly used because their method is computationally efficient and can cover a wide dynamic range. 9 The contribution from all grids is

10. Extracting Feature by Entropy-Based Fractal Dimension Estimation(EBFDE) Our entropy-based fractal dimension estimation (EBFDE) method can further capture the information about randomness of pixels. 10 So the total contribution from all grids is

11. Combination of Two Fractal Dimension Texture Features Here, we allow a small portion of overlapping between two neighboring sub-ranges because there is no clear cut between two sub-ranges reflecting different self-similarity properties. 11

12. Example 12

13. Classification Methods The first classification technique used in this paper for automatic Gleason grading is Bayesian classifier. The second classifier used in this paper for automatic Gleason grading is K-NN which is well-known among all nonparametric classifiers. The third classification technique used in this paper for grading carcinoma prostate images is the SVM method. 13

14. Estimation for Accuracy of Classification Correct classification rate(CCR) 14 Leave-one-out (LOO) and k-fold cross-validation are two popular error estimation procedures to reduce bias in machine learning and testing when sample size is small.

15. Feature Selection Feature selection (FS) is a problem of deciding an optimal subset of features based on some selecting algorithm. The sequential floating forward selection (SFFS) method is very effective in selecting an optimal subset of features. 15

16. Image Acquisition There were 205 pathological images with resolution 512*384 pixels. in Class-1 (Grade-1 and Grade-2): 50 images. in Class-2 (Grade-3): 72 images. in Class-3(Grade-4): 31 images. in Class-4 (Grade-5): 52 images. Images were commonly analyzed by a group of experienced pathologists. 16

17. Feature Sets Used for Comparison In this research, we use the feature sets derived from multiwavelets, Gabor filters, and GLCM methods to compare with our FD-based feature set and demonstrate the superiority of our approach over other methods. 17

18. Performance of FD-Based Feature set(1/2) 18

19. Performance of FD-Based Feature set(2/2) 19

20. Comparison of CCR Using Classifiers Without Feature Selection(1/3) 20

21. Comparison of CCR Using Classifiers Without Feature Selection(2/3) 21

22. Comparison of CCR Using Classifiers Without Feature Selection(3/3) 22

23. Comparison of CCE Using Classifiers With Feature Selection(1/3) 23

24. Comparison of CCE Using Classifiers With Feature Selection(2/3) 24

25. Comparison of CCE Using Classifiers With Feature Selection(3/3) 25

26. Discussion And Conclusion Experimental results demonstrated that the FD- based feature set proposed in this paper can provide very useful information for classifying pathological prostate images into four classes in Gleason grading system. We successfully propose a fractal dimension feature set of very small size to grade prostate images effectively. 26

27. SVM 27