1. of 39 Vector Field Analysis for Segmentation of the Ultrasound Images of Breast Cancer Stanislav S. Makhanov School of Information and Computer Technology, Sirindhorn International Institute of Technology, Thammasat UniversityThailand

2. of 39 Biomedical Engineering Unit, Sirindhorn International Institute of Technology Prof. Dr. Stanislav Makhanov,Unit Head, medical image processing, CNC machining of medical implants Assoc. Prof. Dr. Bunyarit Uyyanonvara medical image processing, retinal image analysis Assoc. Prof. Dr. Waree Kongprawechnon, control engineering Assoc. Prof. Dr. Siriwan Suebnukarn, Department of Dentistry, TU, dentistry, virtual reality Dr. Itthisek Nilkhamhang control engineering, robotics Assoc. Prof. Dr. Utairat Chaumrattanakul, Department of Radiology: ultrasound diagnostics Asst. Prof. Toshiaki Kondo medical image processing Asst. Prof. Dr. Pakinee Aimmanee retinal image processing

3. of 39 “Breast cancer is the chief cause of natural death among women in Thailand. In 2012 alone, Thailand saw 13,184 new cases, leading to 4,665 deaths. Bluntly put, there is roughly one death from breast cancer every two hours” (The Nation, October,2013) Stage5-year Survival Rate 0100% I II93% III72% IV22% Breast cancer survival rates. The US National Cancer Institute’s Database Breast cancer

4. of 39 Ultrasound imaging for breast cancer diagnostics One of the most practical and rapid method of obtaining breast cancer images Relatively inexpensive technology with the benefits of portability Non invasive, real-time and free of harmful radiation

5. of 39 Input: the US image Output: the contour of the tumor Segmentation of Ultrasound Images

6. of 39 Snakes (active contours) Active contours or snakes are closed evolving curves designed for segmentation of complex shaped 2D objects object

7. of 39 resists stretching resists bending (corners) The external force moves the snake towards the object. The force is based on the image features such as edges elasticitystiffness 7 The curve is treated as an elastic rubber band characterized by elasticity and stiffness Mathematical Model

8. of 39 Snake Based Segmentation

9. of 39 attracts the snake to the intensity I desired attracts the snake to large gradients of the image gray level attracts the snake to the edges External force

10. of 39 Some typical sources of the errors

11. of 39 Small capture range Gradient vector field Edge map Gradient field Small capture range

12. of 39 Noise, shadows false boundaries ground truth snake gets distracted by the noise Cancer tumor in an ultrasound image Noise

13. of 39 Inappropriate parameters It is often hard to find appropriate parameters of the snake ?? Wrong parameters

14. of 39 The snake evolves too fast or too slow. It stops at the false boundary or does not stop at all. Slow convergence It is hard to define an appropriate convergence criterion. When enough is enough ? Breast cancer segmentation from an ultrasound image No convergence

15. of 39 Poor initialization Inappropriate seed

16. of 39 Concave objects Wrong result

17. of 39 Snake can not automatically split into several snakes to detect multiple objects Multiple objects Wrong result

18. of 39 Embedded objects Snakes can not automatically detect the structure of embedded objects 1 2 3 4 1 2 4 3 Wrong result

19. of 39 Gradient Vector Flow: Mathematical model Diffusion, smooth and extends the vector field Stopping term Gradient vector flow field Conventional Improved diffusion

20. of 39 Are able to repel or attract each other, split, merge, and disappear as necessary. They are also capable of producing offspring anti-snakes. Multiple Cooperating High Order Snakes The snake points repel each other when they are close enough Contracting snake Expanding snake The two snakes capture the object

21. of 39 Multiple snakes

22. of 39 Vector field analysis

23. of 39 Phase portrait analysis Noise or boundary noise background boundary

24. of 39 Force Field Analysis: Direction score Rotating window Measuring the deviation from anti-parallel positions

25. of 39 Direction score

26. of 39 Direction score as an edge detector The best conventional edge detector Direction score

27. of 39 Gradient Vector Flow with the Direction Score

28. of 39 The edge magnitude score OriginalPreprocessedEdge map Clustering 1 2 3 4 Magnitude score 1 2 34

29. of 39 Gradient Vector Flow with the edge magnitude

30. of 39 Multi-feature diffusion orientation (force field analysis) Magnitude (clustering) Weak edge Strong edge Weak edge shadows no edge

31. of 39 Reference methods Classical GGVF with manual training Criterion/ Method No preprocessing GGVF + best K 0.09 Preprocessing GGVF manualMulti-feature vector flow Best K=0.13 Acceptable range: K=0.12-0.134 TP 24.916588.005988.005091.0236 H1H1 32.02787.08817.35315.6857 H2H2 23.96051.39351.39390.6766 H3H3 13.48370.78410.78430.3808 Snake convergen ce yes GGVF convergen ce explicit scheme no yes TP – true positives, H 1 - max Hausdorff distance, H 2 - average Hausdorff distance, H 3 relative to the contour length Hausdorff distance x 1000

32. of 39 Reference methods Poisson gradient vector flow C.Y. Hsu, C.Y. Liu, C.M. Chen, Automatic segmentation of liver PET images, Computerized Medical Imaging and Graphics, vol. 32, 2008, pp. 601–610. C.Y. Hsu, K.F. Wang, H.C. Wang, K.K. Tseng, Automatic extraction of face contours in images and videos, Future Generation Computer Systems, vol 28, 2012, pp. 322–335 (M2012) Mixed noise vector flow O. Ghita, P. F.Whelan, A new GVF-based image enhancement formulation for use in the presence of mixed noise, Pattern Recognition, vol 43, 2010, pp. 2646–2658 (M2010) B. Li, S.T. Acton (2007) Active contour external force using vector field convolution for image Segmentation, IEEE Transactions on Inage Processing, vol 16(8), pp 2096-2106 (M2007) Convolution vector flow J. Cheng, S.W. Foo, Dynamic directional gradient vector flow for snakes, IEEE Transactions on Image Processing, vol 15( 6), 2006, pp. 1563-1571 Directional gradient vector flow

33. of 39 Testing on synthetic tumors

34. of 39 Testing against recent methods on synthetic tumors Noise/ Criterion/Method M2010 M2012 M2007 Multi-feature diffusion PSNR=24db % images, TP>80% 100% 97.4722 6.3956 2.0571 100% 99.83% 5.1006 1.6657 100% 99.81% 5.6978 2.1047 100% 99.89% 4.5927 1.5827 Average TP Average H 1 Average H 2 PSNR=17db % images, TP>80% 52.63% 85.71% 11.9572 3.4501 78.90% 88.99% 9.9474 2.7706 78.95% 92.24% 8.6094 2.1379 100% 96.17% 4.7755 1.6064 Average TP Average H 1 Average H 2 PSNR=15db % images, TP>80% NA NA 52.63% 86.23% 14.9318 4.6069 84.21% 91.83% 8.9325 3.3364 Average TP Average H 1 Average H 2 # of training parameters 3 3 3 none

35. of 39 Testing recent methods on the US image database onlinemedicalimages.com

36. of 39 Multi feature diffusion vs. recent methods Criterion/Method M2010 M2012 M2007 Multi-feature diffusion % images, TP>80% 64% 87.581 1.5860 2.9513 64% 89.86% 8.1425 2.3916 71% 89.07% 8.2573 2.4074 79% 92.06% 5.1706 1.5099 Average TP Average H 1 Average H 2 # of training parameters 3 3 3 none

37. of 39 Multi-feature vs. the standard methods Canny edge detection Direction score Magnitude score

38. of 39 Multi-feature vs. the standard methods Original Ground truth M2011 M2012M2007 Multi-feature

39. of 39 Other Applications

40. of 39 Original imageExtraction results Extracting roads from satellite Images

41. of 39 Defeating Digital CAPTCHAS 78% recognition rate

42. of 39 Thank you