Room: AAPM ePoster Library
Prostate bed segmentation in planning CT image is essential to the post-prostatectomy radiotherapy. However, due to the non-contrast boundaries and highly variable appearance, the delineation of the prostate bed, a treatment target, is challenging and typically carried out by the physicians. In this work, we built an atlas model to segment the prostate bed and surrounding organs-at-risk (OAR) automatically.
186 post-prostatectomy cases from the year 2009 to 2019 were collected. A five-fold cross-validation strategy was used in this study. Specifically, the collected cases were randomly divided into five folds. Each fold was used as a testing group alternatively, while the rest four folds were used for building up the atlas model with MIM software. The testing case was first compared to the atlas to find out four best-matched atlas cases, which were then registered to the testing case by a deformable registration. The contours from these matched atlas cases were transformed into the testing case. Finally, a majority voting algorithm combined the four transformed contours into one to create the final segmentation result. The experimental results were evaluated by the dice similarity coefficient (DSC) and average symmetric surface distance (ASD).
The global DSC of the atlas-generated contours of the prostate bed, the bladder, and the rectum are 64.21±11.88%, 64.07±17.48%, and 61.75±11.54%, respectively, and the global ASD are 4.81±11.40mm, 7.65±15.27mm, and 5.85±10.99mm, respectively. The average segmentation time for the ABS workflow is around 3 minutes per CT image.
We establish a fully automatic workflow for prostate bed segmentation by building an atlas-based model. The proposed approach can segment the prostate bed and the OARs efficiently and reproducibly. Although the result still needs further improvement for full clinical use, it can serve as a good initial contour to save the manual contouring time.
Funding Support, Disclosures, and Conflict of Interest: The research is in part supported by NIH grant 1R01CA206100.
Targeted Radiotherapy, Segmentation