1. 1 SEGMENTATION OF BREAST TUMOR IN THREE- DIMENSIONAL ULTRASOUND IMAGES USING THREE- DIMENSIONAL DISCRETE ACTIVE CONTOUR MODEL Ultrasound in Med. & Biol., Vol. 29, No. 11, 2003 資訊所 P76994458 石慧萱 2011.01.06 HCI Final Report

2. 2 Outline Introduction Materials And Methods –Review –Proposed 3-D Segmentation Method Result Conclusion

3. 3 Introduction

4. 4 Clinically, ultrasound is widely used for tumor detection. Although the artificial recognition method is good, manual handling several hundreds of images in a 3-D data set is a time-consuming process. Precise 3-D segmentation approach can provide an accurate evaluation of the tumor volume and solid tumor shape

5. 5 Introduction In this paper, we apply three-dimensional active contour model to a 3-D ultrasonic data file for segmenting of the breast tumor However, there is emphasis on these 3-D techniques that they do not consist of a series of 2-D techniques. When they work, they will consider the horizontal, vertical and depth directions at the same time.

6. 6 Introduction The ultrasound imaging has some characteristics including noise, speckle and tissue-related textures. The conventional edge-based image segmentation and region-based segmentation are not suited to find tumor boundaries. Stick is a boundary detection approach based on an image enhancement technique.

7. 7 Introduction We propose a new 3-D active contour model based on the traditional 2-D snake to find the 3-D shape in a 3-D US data set. The 3-D active contour model will make the initial contour approach to the real contour of the tumor.

8. 8 Intoduction Extract 3D volume data 3D stick procedure Apply the auto threshold method to obtain the binary image Apply the 3D morphological process (close & opening) to obtain the initial 3D shape of the tumor 3D snake procedure Final 3D shape of the tumor Initial Shape Finding Final Shape Deformation Flow Chart of the Proposed Method

9. 9 Introduction The use of these 3-D techniques not only segments the 3-D shape but also obtains the volume of the tumor. The volume of the tumor calculated by the proposed method will be compared with the volume calculated by the software with the physician’s manually drawn shape.

10. 10 Review

11. 11 3-D Breast US In this work, the 3-D sonography was performed using a Voluson 530D (Kretz Technik, Austria)scanner and a Voluson small part transducer S-VNW5 to 10. A linear-array transducer with a frequency 5 to 10 MHz, a scan width of 40 mm and the sweep angle of 20° to 30°

12. 12 3-D Breast US These 3-D volume file should be saved in Cartesian coordinates by the Voluson 530D or 3-D View 2000 program Three planes of a volume data set could be obtained by using our developed program –Longitudinal –Transverse –Coronal planes

13. 13 Reviews of Stick Using line segments (also called sticks) in different angular orientations as a template Large-scale linear features line up with sticks of adequately short length, while speckle does not if the sticks are long enough Select the most suitable orientation at each point To reduce speckle and improve edge information in ultrasonic images

14. 14 Reviews of Stick A square N x N area in an image 2N - 2 short lines that pass through the central pixel,with each line including exactly N pixels Calculate the average of pixel values along the line, select the maximum average as the value of center pixel 5x5 area 2*5 -2=8 組 stick

15. 15 Reviews of Snake An established method for the contour extraction and image interpretation The contour is represented as a set of vertices that move in response to: –Internal forces : derived from the properties of the shape of the contour, minimize the local contour curvature –External forces : derived from the image features, make the model follow a path of low energy through the external energy distribution

16. 16 Reviews of Snake The deformation process: The forces acting on a vertex x i are internal, external and damping forces w int, w ext and w damp are the weighting factors, v i is the velocity of vertex x i

17. 17 Proposed 3-D segmentation method

18. 18 Flow Chart Extract 3D volume data 3D stick procedure Apply the auto threshold method to obtain the binary image Apply the 3D morphological process (close & opening) to obtain the initial 3D shape of the tumor 3D snake procedure Final 3D shape of the tumor Initial Shape Finding Final Shape Deformation Flow Chart of the Proposed Method

19. 19 3-D Stick Two kinds of stick length, five and seven, are used in the proposed 3-D stick Calculated based on the symmetry Defined based on how many stick elements are included in one frame

20. 20 3-D Stick Length = 5, b (1 – 3 -1 ) x z y

21. 21 3-D Stick Two sample forms in (a) category a, and (b) category b of the stick with length five Two sample forms in (a) category a, (b) category b and (c) category c of the stick with length seven.

22. 22 Flow Chart Extract 3D volume data 3D stick procedure Apply the auto threshold method to obtain the binary image Apply the 3D morphological process (close & opening) to obtain the initial 3D shape of the tumor 3D snake procedure Final 3D shape of the tumor Initial Shape Finding Final Shape Deformation Flow Chart of the Proposed Method

23. 23 Automatic Thresholding Minimize the sum-of-square errors SSE(t) between the gray values and the mean values and hi is the number of pixels with gray level i.

24. 24 Flow Chart Extract 3D volume data 3D stick procedure Apply the auto threshold method to obtain the binary image Apply the 3D morphological process (close & opening) to obtain the initial 3D shape of the tumor 3D snake procedure Final 3D shape of the tumor Initial Shape Finding Final Shape Deformation Flow Chart of the Proposed Method

25. 25 3D Morphology Use morphologic techniques – Opening : erosion + dilation – Closing : dilation + erosion Extend the morphologic operations from 2- D to 3-D use a 5x5x5 cube in the 3-D morphologic filtering Use a 5x5 square for morphologic edge detection

26. 26 Flow Chart Extract 3D volume data 3D stick procedure Apply the auto threshold method to obtain the binary image Apply the 3D morphological process (close & opening) to obtain the initial 3D shape of the tumor 3D snake procedure Final 3D shape of the tumor Initial Shape Finding Final Shape Deformation Flow Chart of the Proposed Method

27. 27 3D Snake- Internal Force The proposed 3-D snake model modifies the internal forces and the external forces to the form of 3-D Internal forces are related with the local contour curvature Besides the curvatures of the horizontal and vertical directions, the internal forces also calculate the curvature of the depth direction

28. 28 3D Snake- Internal Force The local curvature c i at x i : XiXi

29. 29 3D Snake- Internal Force The locally tangential unit vector at a vertex X i : The locally tangential unit vector at a vertex XiXi

30. 30 3D Snake- Internal Force The unit vector in the local radial orientation is derived from by a rotation over π/2 radians: XiXi

31. 31 3D Snake- Internal Force The internal force f int,i can be defined as : (8) (9)

32. 32 3D Snake- External Force In the conventional snake model, the external force is usually calculated from the gradient of the original image The proposed external force is calculated from the texture information The Gaussian blurring is applied for better results The original image feature is also used to retain the boundary information

33. 33 3D Snake- External Force I(x,y,z): the original image T(x,y,z): the texture image after applying the texture analysis to I Texture analysis: calculate the average variance of the voxel, centered at point P(x,y,z), with the size of 3x3x3 pixels T GB (x,y,z) : the Gaussian blurring operator is applied to the texture image T(x,y,z) -I’ and –T GB ’ : derived from normalizing -I and -T GB into [0,1].

34. 34 3D Snake- External Force The external energy distribution E is defined as the sum of –I’ and –T’ GB, and the external force f ext can be calculated from E

35. 35 3D Snake-Deformation Process t : the step at a particular time t ∆t: the incremental time

36. 36 3D Snake- Deformation Process The distance between two vertices will change constantly Periodically resampling the model along its path: – The segment length > upper bound – The segment length < lower bound

37. 37 Calculation of Volume of Tumor The region-growing method can be used to fill up the region encircled by the final shape Volume: The number of voxels in the filled regions

38. 38 Result

39. 39 Data Benign Malignant

40. 40 Resulting Images (a) The original image. (b) The equalized image of (a) (c) The stick image of (b) (d) The image after automatic threshold of (c) (e) The image after morphologic process of (d) (f) The contour of the tumor derived from (e) (g) The result of the snake

41. 41 Resulting Images (a) The original image. (b) The equalized image of (a) (c) The stick image of (b) (d) The image after automatic threshold of (c) (e) The image after morphologic process of (d) (f) The contour of the tumor derived from (e) (g) The result of the snake

42. 42 Resulting Images (a) The original image. (b) The equalized image of (a) (c) The stick image of (b) (d) The image after automatic threshold of (c) (e) The image after morphologic process of (d) (f) The contour of the tumor derived from (e) (g) The result of the snake

43. 43 Resulting Images

44. 44 Expert Annotation The VOCAL TM software contained in the 3-D VIEW TM is a tool for calculating the volume data as well as the geometric surface information in a well-defined border lesion

45. 45 Comparison The average match rate is about 95%

46. 46 Performance Morphology

47. 47 Conclusion

48. 48 Conclusion The 3-D techniques, including the 3-D stick, 3-D morphologic process and 3-D snake model, are first used and the outcome is similar to that of the VOCAL TM software with the physician’s manual adjustment The extension of the traditional 2-D segmentation approach and the use of the depth information between image slices make the segmentation result more precise The introduced 3-D ultrasound segmentation method not only provides the accurate segmentation of the tumor shape, but also evaluates the volume of the tumor.

49. 49 Conclusion Physician does not need to spend a lot of time manually drawing the initial contour in each ultrasonic image of the tumor The obtained final 3-D shape of the tumor can be separated into the contours in image slices and these contours can be treated as the 2-D segmentation result of each ultrasonic image of the tumor The volume information of the tumor can be used to trace the variant state of the tumor in clinical applications