1 / 65

Cell-Based Image segmentation for 2D and 2D series ultrasound images 以區域單元為基礎之超音波影像與超音波序列影像分割

MIRAGE Lab. NTU BME. Cell-Based Image segmentation for 2D and 2D series ultrasound images 以區域單元為基礎之超音波影像與超音波序列影像分割. Student: Cheng, Jie-Zhi Thesis Adviser Dr. Chen, Chung-Ming 1 Dr. Chou, Yi-Hong 2 1 Institute of Biomedical Engineering, NTU 2 Department of Radiology, TVGH.

Télécharger la présentation

Cell-Based Image segmentation for 2D and 2D series ultrasound images 以區域單元為基礎之超音波影像與超音波序列影像分割

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. MIRAGE Lab NTU BME Cell-Based Image segmentation for 2D and 2D series ultrasound images以區域單元為基礎之超音波影像與超音波序列影像分割 Student: Cheng, Jie-Zhi Thesis Adviser Dr. Chen, Chung-Ming1 Dr. Chou, Yi-Hong2 1Institute of Biomedical Engineering, NTU2Department of Radiology, TVGH

  2. MIRAGE Lab NTU BME Outline • 2D image segmentation– Augmented Cell Competition Algorithm • 2D series image segmentation– Cell-Based Two Region Competition with MAP Framework (C2RC-MAP) • Performance Analysis • Summary

  3. weak edge MIRAGE Lab NTU BME Difficulties in 2D Segmentation • Weak edges • Speckle Noises • Shadowing effect • Artifacts t130 t90 Initial level set

  4. MIRAGE Lab NTU BME Difficulties in 2D Segmentation • Complicated Echogenicity

  5. MIRAGE Lab NTU BME Augmented Cell Competition (ACCOMP) algorithm • A nearly automatic, data-driven, segmentation algorithm • Being composed of two phases • Image/ROI Partition: Cell Competition Algorithm • Edge Grouping: Cell-Based Graph-Searching Algorithm • Five best contours are suggested

  6. lesion boundary prominent component MIRAGE Lab NTU BME Motivation of Two-Phase 1.Object of Interest could possibly constitute of Several Prominent Components

  7. MIRAGE Lab NTU BME Motivation of Two-Phase 2. Edge grouping by exploring the bilateral consistence along the contour

  8. MIRAGE Lab NTU BME Partition Phase/Cell Competition Algorithm- The Basic Ideas • ROI is first decomposed into cells, each of which is a homogeneous area • Cell-based deformation: Only the cell boundaries are considered as the candidate positions for deformation • Cell Competition: iteratively split/merge the cells into prominent components

  9. MIRAGE Lab NTU BME Cell Generation Gradient image Simulated image Smoothed image 2nd watershed 1st watershed

  10. MIRAGE Lab NTU BME Benefits of Cell-Based • More Efficient-- Less Search Space • Concrete Structural Region and Edge Information-- facilitate the integration of region and edge clues • Statistically More Robust to Noise-- cell is a region of pixels with similar intensities

  11. Action A Action B Action C MIRAGE Lab NTU BME Cell Competition: Three Action Types

  12. MIRAGE Lab NTU BME Cost Function for cell competition in the ith iteration Characterize Homogeneity Characterize Boundary Salience

  13. MIRAGE Lab NTU BME An Example

  14. MIRAGE Lab NTU BME Edge Grouping Phase/Cell-Based Graph-Searching Algorithm- The Basic Ideas • Select And Group the edge segments in the prominent component tessellation • Propose five boundary candidates • Bilateral Consistence Exploration • Implement in Graph Traversing Scheme

  15. ce1 ce3 ce2 ce4 ce12 ce6 ce5 ce11 ce10 ce13 ce7 ce9 ce8 ce3 ce1 ce4 ce2 ce5 ce12 ce6 ce10 ce11 ce7 ce13 ce9 ce8 MIRAGE Lab NTU BME Cell-based Graph-Search Algorithm- construct c-graph Cell Edge Prominent Component Tessellation c-graph

  16. MIRAGE Lab NTU BME Cell-based Graph-Search Algorithm- generate an initial guess • Combine the result of a conventional segmentation algorithm and the region boundaries to get an initial guess • Rough Outline from Region Competition

  17. MIRAGE Lab NTU BME Cell-based Graph-Search- search potential boundaries • Boundary of object of interest should be1. closed2. not self-intersected • Correspond to a cycle in c-graph • Use Depth-First Search scheme • Exploration of Bilateral Consistence within the DFS scheme • See an example

  18. MIRAGE Lab NTU BME Cell-based Graph-Search- Select the best boundaries • Five best boundaries are suggested according to five cost functions • The cost function is a function of the gray level distribution of the banding area along the boundary.

  19. MIRAGE Lab NTU BME Cell-based Graph-Search- Select the best boundaries • C1: Overall absolute difference of bilateral mean intensities: • C2: Continuity of mean gradient: • C3: Overall edge strength: maximize minimize maximize

  20. MIRAGE Lab NTU BME Cell-based Graph-Search- Select the best boundaries • C4: Continuity of mean intensities: • C5: Sum of the continuity of mean intensities and the negative overall edge strength: minimize maximize

  21. MIRAGE Lab NTU BME Five Suggested BoundariesDemonstration

  22. MIRAGE Lab NTU BME Every Significant Steps of Augmented Cell Competition Algorithm– Demonstration(1) • 1st watershed 2nd watershed Prominent Components Rough outline 1st candidate • 2nd candidate 3rd candidate 4th candidate 5th candidate manual delineation Note: The edge segment emphasized byellipsoid in Prominent Component is the initial edge for the second phase

  23. MIRAGE Lab NTU BME Every Significant Steps of Augmented Cell Competition Algorithm– Demonstration(2) • 1st watershed 2nd watershed Prominent Components Rough outline 1st candidate • 2nd candidate 3rd candidate 4th candidate 5th candidate manual delineation Note: The edge segment emphasized byellipsoid in Prominent Component is the initial edge for the second phase

  24. MIRAGE Lab NTU BME ACCOMP Performance Analysis • Each image was manually delineated by four observers • Four contours were derived by ACCOMP algorithm for each images • Three assessments were carried out for each series • The ACCOMP is tested upon 300 breast sonograms, including 165 carcinomas and 135 fibroadenomas

  25. MIRAGE Lab NTU BME ACCOMP Performance Analysis • First assessment : check if the computer-to-observer distance is less than the maximum interobserver distance

  26. MIRAGE Lab NTU BME ACCOMP Performance Analysis • Second assessment : tests if there is a significant difference among the four sets of computer-generated boundaries of a series group with respect to the average manually delineated boundaries. • The averaged distances of the four sets to the averaged manual delineations are 3.36±2.36, 3.41±2.32, 3.49±2.43, and 3.40±2.33 respectively. • Friedman test was used and The p value is 0.54.

  27. MIRAGE Lab NTU BME ACCOMP Performance Analysis • Third assessment : computes the Pearson’s correlation of the lesion areas defined by the computer-generated boundaries and manually-delineated boundaries • The Pearson’s correlations were all higher than 0.98.

  28. MIRAGE Lab NTU BME Summaries • The advantage of two phases:1. Partition phase: further scale down the search space with meaningful structure2. Edge Grouping: bypass the complicated echogenicity • The ACCOMP algorithm is a model-free image segmentation method in which training scheme is not necessary. • The ACCOMP algorithm is capable of delineating highly winding boundaries and dealing irregular echo pattern within the boundary.

  29. MIRAGE Lab NTU BME More Demonstration

  30. MIRAGE Lab NTU BME More Demonstration

  31. MIRAGE Lab NTU BME More Demonstration

  32. MIRAGE Lab NTU BME More Demonstration

  33. MIRAGE Lab NTU BME Related Works • C. M. Chen, Y. H. Chou, C. S. K. Chen, J. Z. Cheng, Y. F. Ou, F. C. Yeh, K. W. Chen. "Cell Competition Algorithm: A New Segmentation Algorithm for Multiple Objects with Irregular Boundaries in Ultrasound Images," Ultrasound in Medicine and Biology, vol. 31, no. 12, pp. 1647-1664, 2005. (SCI; 2005 Impact Factor 2.221) • C. M. Chen, J. Z. Cheng, Y. H. Chou. “ACCOMP— Augmented Cell Competition Algorithm for Delineating Boundaries of Objects of Interest in Sonography,” Technical Report, Institute of Biomedical Engineering, National Taiwan University.

  34. MIRAGE Lab NTU BME Segmentation on 2D series • Why not manual delineation?1. Less coherence2. Tedious

  35. MIRAGE Lab NTU BME Segmentation on 2D series • Knowledge-based • Model-based • General approach

  36. MIRAGE Lab NTU BME Segmentation on 2D series- Knowledge-based • incorporating the intrinsic shape properties of the object of interest as the prior knowledge into the segmentation algorithms • For example: Dydenko et al. (2006) march the frontier of a level set constrained by the shape of myocardium.

  37. MIRAGE Lab NTU BME Segmentation on 2D series- model-based • find the 2D boundaries based on the mathematical shape models describing the common characteristics of the objects of interest or the shape models constructed from the training data. • For example: Bosch et al. (2002) and Xie et al. (2005) represent the boundaries of the LV endocardial contours and the kidneys, respectively, as a linear combination of the mean shape and a set of eigenshapes computed from the training data.

  38. MIRAGE Lab NTU BME Segmentation on 2D series- general approach • Without using prior shape knowledge or model information • For example: the contour points within a slice and between adjacent slices are modeled as a Markov random field (MRF) in Hass et al. (2000) to control the smoothness of the derived IVUS boundaries.

  39. MIRAGE Lab NTU BME C2RC-MAP Algorithm • Based on two essential ideas • cell-based two-region competition • cell-based MAP framework • Starting from the boundary derived for the initial slice, the boundary of previous slice is used as a reference for the current slice. • Propagate the derived contour as a reference contour for the next slice

  40. MIRAGE Lab NTU BME Segmentation on Initial Slice • Augmented Cell Competition Algorithm • A nearly automatic segmentation algorithm composed of • Cell Competition algorithm • Cell-based graph-search algorithm • Five best contours are suggested

  41. MIRAGE Lab NTU BME Cell-based Two-region Competition for Each Slice In each iteration, the object- and background-regions compete the cells along the region contour and move the cell that results in the maximal improvement of the cost function.

  42. MIRAGE Lab NTU BME Cell-based MAP Framework • I : the set of mean gray levels of the cells in the ROI of the current slice • C : the regional contour separating the object-and background-regions • L : the labeling vector of the cells in the ROI Prior model Region appearance probability model Contour model

  43. MIRAGE Lab NTU BME Roles of Three Models • Region Appearance Probability Model : quantifies the cell-based textural property for the object- and background-regions • Contour Model : emphasizes on the edge properties of the regional contour • Prior Model : regularizes the shape complexity of the object-region with a preference for a smooth regional contour to a protruding one

  44. MIRAGE Lab NTU BME Region Appearance Probability Model • describes the conditional joint probability of the mean gray levels of the cells in both regions • Conditioned on L and C, the mean gray level of a cell is assumed to be normally distributed

  45. MIRAGE Lab NTU BME Contour Model • a likelihood function that characterizes the salience of the regional contour and the coherence of the regional contour to the reference contour given an instance of the cell labeling • Two terms:1. Coherence: Securing the contextual coherence2. Salience: Characterize contour salience

  46. MIRAGE Lab NTU BME Contour Model

  47. MIRAGE Lab NTU BME Prior Model • The labeling of each cell is modeled as Gibbs distribution • The more neighboring cells have the same label, the smaller U(L) would be, and hence the larger p(L) would be

  48. MIRAGE Lab NTU BME Effect of Prior Model Avoidance of Chaining

  49. MIRAGE Lab NTU BME Optimized by EM Algorithm • E-step: Given model parameters, find the single-cell migration from one region to the other that results in the largest positive increment of the log-posterior: • M-step: Given the new cell labeling and the new regional contour, compute the new model parameters for next E-step

  50. MIRAGE Lab NTU BME Results original manual level set C2RC

More Related