Fifty-eighth annual meeting of the american association of physicists in medicine
TH-CD-206-03: Evaluation of a New Method to Improve Atlas Based Segmentation for the Bladder and Rectum
Atlas segmentation for the bladder and rectum pose particular challenges due to variability in bladder filling and variability in rectal contents including air and fecal matter. Previously we evaluated atlas segmentation for high risk prostate cancer including the bladder and rectum. In the current study we sought to further improve automated segmentation results by utilizing a customized post-processing workflow for the bladder and rectum.
Pelvic CT scans for 32 subjects with prostate cancer were chosen and the bladder and rectum were first segmented using the Multi-5 atlas-based approach described in Pirozzi et al (IJROBP 2012). The post-processing workflow for the bladder limited the HU range to include only water and soft tissue density voxels (Multi-5 PP). For the rectum, air in the image was masked to soft tissue density and an atlas masked with the same approach was used to segment the rectum. Next, the rectal contours were restricted to exclude voxels in the fat HU range (Masked Multi-5 PP). Multi-5, Multi-5 PP, and Masked Multi-5 PP were compared to manually defined contours using Hausdorff distance and the Dice Similarity Index (DSI).
The average dice and Hausdorff distance for the bladder for Multi-5 and Multi-5 PP was 0.83, 0.46 and 0.86, 0.39 respectively, significantly increasing dice while decreasing Hausdorff distance (p < 0.0005). Multi-5 PP for the rectum improved significantly over Multi-5 alone with DSI and Hausdorff distances of 0.79, 0.31 and 0.68, 0.43 respectively (p-value < 0.0005).
Customized workflows applying post-processing steps to atlas-segmented bladders and rectal masking significantly increase the accuracy of contours when compared to manual contours. These automated steps should be incorporated into the atlas based segmentation workflow of the bladder and rectum.
Aaron Nelson is a part owner and employee of MIM Software Inc.; Sara Pirozzi is an employee of MIM Software Inc.; Stephanie Chung is an employee of MIM Software Inc.