Our segmentation results are visually satisfactory except at organ junctions. We've tried to mitigate these effects using "odd-slice" detection in pre and post processing. As a whole, the system operates faster than slice-by-slice hand segmentation and has viability for "human-in-the-loop" systems.
Future Directions:
Inter-slice edge interpolation to strengthen weak edges, and/or fill in edges where they are suggested. In addition, when a region encounters an edge voxel, we could tighten the growing threshold for the edge median blocks.
Post processing steps to compare the number of voxels to known heuristics for organ size, and extended these heuristics to find low-confidence voxel clumps to remove.
Since kidney cross-sections are generally ellipsoidal, a Hough transform could be used to search for these ellipsoids. This approach applied to a local region around questionable slices may allow a kidney to be found even its adjacent organ edge is non-existent.
References
[1] T. S. Newman, N. Tang, S. Bacharach, P. Choyke, "A Volumetric Technique for Diagnosis and Surgical Planning Lower Torso CT Images." 1015-4651/96 Proceedings of ICPR 1996.
[2] B. Tsagaan, A. Shimizu, H. Kobatake, K. Miyakawa, Y. Hanzawa, "Segmentation of Kidney Using a Deformable Model," 7803-6725 Proceedings of the IEEE January 200.1
|
|
|