Learning Distance Transform for Boundary Detection and Deformable Segmentation in CT Prostate Images.

摘要: Segmenting the prostate from CT images is a critical step in radiotherapy planning for cancer. The segmentation accuracy could largely affect efficacy of radiation treatment. However, due to touching boundaries with bladder and rectum, boundary often ambiguous hard recognize, which leads inconsistent manual delineations across different clinicians. In this paper, we propose learning-based approach detection deformable prostate. Our proposed method aims learn distance transform, maps an intensity image into map. To enforce spatial consistency on learned combine our auto-context model iteratively refining estimated After refinement, can be readily detected by finding valley addition, map also used as new external force guiding segmentation. Specifically, automatically segment prostate, integrate level set formulation. Experimental results 73 show that transform more effective than traditional classification-based driving Also, achieve consistent segmentations human raters, accurate existing methods under comparison.