Thesis Advisor: Dr. Mandayam Committee: Dr. Kadlowec and Dr. Polikar

1. Automated evaluation of radiodensities in a digitized mammogram database using local contrast estimation Thesis Advisor: Dr. Mandayam Committee: Dr. Kadlowec and Dr. Polikar Friday, July 23, 2004

2. Outline • Introduction • Objectives of the Thesis • Previous Work • Approach • Results • Conclusions

3. Introduction

4. Breast Cancer

5. Survival Rates Each stage designates the size of the tumor how much it has spread. Stage 0 Cancer: Lobular Carcinoma in Situ (LCIS) Ductal Carcinoma in Situ (DCIS) 20% of all diagnosed cancers

6. Compression Plate Film Holder Pectoral Muscle Film Holder Mammography Procedure CC View MLO View Compression Plate

7. Risk Factors

8. Chest wall Pectoralis muscles Lobules Nipple surface Areola Duct Fatty tissue Skin Breast density Radiodense Tissue Radiolucent Tissue

9. Chest wall Pectoralis muscles Lobules Nipple surface Areola Duct Fatty tissue Skin Radiodensity Radiodense Tissue Radiolucent Tissue

10. Mammographic Density • “……..women who had a breast density of 75% or greater had an almost fivefold increased risk of breast cancer…………” • – Byrne, C, et. al. “Mammographic features and breast cancer risk: effects with time, age, and menopause status,” Journal of the National Cancer Institute, Vol. 87, pp.1622-1629, 1995.

11. Genetic Heritability • “Women with extensive dense breast tissue visible on mammogram have a risk of breast cancer that is 1.8 to 6.0 times that of women of the same age with little or no density.” • “…………….. the percentage of dense tissue on mammography at a given age has high heritability. Because mammographic density is associated with an increased risk of breast cancer, finding the genes responsible for this phenotype could be important for understanding the causes of the disease.” • – Boyd, N.F., et al, “Heritability of mammographic density, a risk factor for breast cancer,” New England Journal of Medicine, Volume 347(12), September 19, 2002, pp. 886-894.

12. Issues • Current methods are still slow and subjective. • Variability still exists between radiologists. • Automated algorithm for fast and objective estimation. • Rowan University.

13. Objectives of this Thesis • Investigate the use of textures for the segmentation of radiodense tissue in a digitized mammogram. • Create an automated algorithm that is able to consistently evaluate digitized mammograms throughout several databases. • Compare the results of the algorithm to a established manual methods, the “Toronto” method as well as previous methods created at Rowan University.

14. Textures Mammogram f1 f2 f3 … fn Region of Mammogram … Texture Description Method Feature 2 Feature 1 Feature … Feature 3 Total Bank of Features Classified Mammogram Un-supervised Clustering Method

15. Automated Algorithm Database 1 Database 2 Automated Process Estimated percentages for database 1 Estimated percentages for database 2 Validation Percentages Compare Compare PERFORMANCES ‘”SHOULD” BE SIMILAR Performance on database 1 Performance on database 2

16. Previous Work • Wolfe’s classification. • “Toronto” method. • Automated techniques. • “Main goal of research conducted at Rowan University”

17. Wolfe’s Classification • N1: The breast is comprised entirely of fat. • P1: The breast has up to 25% nodular densities. • P2: The breast has over 25% nodular mammographic densities. • DY: The breast contains extensive regions of homogeneous mammographic densities.

18. “Toronto” Method Load Image Load Image into Computer into Computer Set Tissue Set Tissue Set Boundary Set Boundary Threshold Threshold Threshold Threshold 1 4096 1 4096 1 4096 1 4096 Count pixels in Count pixels in regions regions 33.3% RD 33.3% RD Display Display Results Results 66.6% RL 66.6% RL 1 4096 1 4096

19. Automated

20. Neyman-Pearson Classifier Distribution 2 (Radiodense) Distribution 1 (Radiolucent) 12,21=22

21. Constrained Neyman-Pearson Classifier • 1 and 2 are means of distributions, 2 is local variance of image • Varies threshold based on the variance of image from pure Bayesian to 2 • Can compensate for brightness of image and classify image radiodensity • Determine  from training data set

22. Spatially Varying CNP Compression Plate CC View Film Holder Less Density Here Less Stress Here More Density Here More Stress Here

23. Compression Compensation Multiple lowpass filtering operations

24. Issues • Most of these algorithms are not fully automated. • Performance is evaluated in just one type of database.

25. Approach • Texture and image processing methods investigated. • Local Contrast Estimation algorithm. • Investigation of previous methods created at Rowan University.

26. Textures • If radiodense and radiolucent tissue exhibit characteristics that are different from each other, texture… • Evaluation of 3 different ‘types’ of methods. Texture description methods

27. Variance 10 images: 5 FCC 5 Harvard Database Evaluation Inter-image and cross-database statistic evaluation Intra-image characteristics evaluation Individual Variance Imaging Variability of texture characteristics

28. Variance Imaging Histogram Regional Variance Imaging Histogram

29. Gabor Filtering Frequency Domain Spatial Domain

30. 50 regional samples for radiodense tissue 50 regional samples for radiolucent tissue averaged frequency profile for radiodense tissue averaged frequency profile for radiolucent tissue 50 frequency profiles for radiodense tissue 50 frequency profiles for radiolucent tissue 2-Dimensional FFT highest region of difference 2-Dimensional FFT Gabor Filtering

31. Co-occurrence 4 2 1 0 2 4 0 0 1 0 6 1 0 0 1 2 0 0 1 1 0 0 1 1 0 2 2 2 2 2 3 3

32. Co-occurrence Energy Entropy Moments Inverse Moments

33. Law’s Texture Energy Measures • Spatial filters based on three simple vectors: • Averaging L = (1,2,1) • Edges E = (-1,0,1) • Spots S = (-1,2,1) • These 3 vectors can be combined to make 25 separate spatial filters.

34. Law’s Texture Filter L5 = [ 1 4 6 4 1 ] E5 = [-1 -2 0 2 1 ] S5 = [-1 0 2 0 -1 ] W5 = [-1 2 0 -2 1 ] R5 = [ 1 -4 6 -4 1 ] • Averaging L = (1,2,1) • Edges E = (-1,0,1) • Spots S = (-1,2,1) 3 Simple Vectors Filtering Result  Energy measuring function  Complements are summed 25 filters All convolution pairs 15 sets of features 1 4 6 4 1 4 16 24 16 4 6 24 36 24 6 4 16 24 16 4 1 4 6 4 1 • [1 2 1]*[1 2 1] = [1 4 6 4 1] 5 Vectors [ 1 4 6 4 1 ] x = All column by row multiplication pairs Image Filtering +energy measure + addition of Complements 25 matrices (filters) Set of 15 features

35. Clustering • Variance Imaging = 1 feature. • Co-occurrence = 4 features. • Law’s Energy Measure = 15 features. Supervised Learning Techniques are not viable because of the vast variation texture characteristics!!!

36. k-means begininitializen, c, µ1, µ2 …µc doclassify n samples according to nearest µi recompute µi untilno change in µi returnµ1, µ2 …µc end

37. Image Processing Techniques for pre-processing & evaluation • Non-linear Transformations. • Gray level connectivity.

38. Non-Linear Transformation = =

39. Gray Level Connectiveness Both 50% of dark pixels and 50% bright pixels

40. Gray Level Connectiveness • Classify the lowest set of pixels .1 gray values away from each other as radiodense. • Afterwards, all regions were analyzed for connectiveness by classifying regions as connected as long as they were within .1 gray values of each other.

41. Variance Typical Harvard Typical FCCC

42. Variance

43. Variance

44. Gabor Filters Typical Harvard Typical FCCC

45. Co-occurrence Matrix Results of clustering Expected result

46. Co-occurrence Matrix Expected result Results of clustering

47. Law’s Energy

48. Non-linear Transformation

49. Non-linear Transformation

50. Connectiveness