(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purpose: Automatic prostate bed target segmentation based on cone-beam CT (CBCT)-imaging is important for adaptive therapy but extremely challenging. CBCT images present low soft-tissue contrast, and a prostate bed target is an inferred geometry with large appearance variations. We hypothesized that incorporating contouring on planning CT as prior would enhance deep learning model performance for CBCT-based prostate bed segmentation.
Methods: 17 prostate cancer patients who received 5-fraction stereotactic body radiotherapy after radical prostatectomy were included. Each patient had one planning CT image and five daily set-up CBCT images acquired. Prostate bed on CT (PB?CT?) and CBCT (PB?CBCT?) images were contoured by an experienced radiation oncologist. We built a multi-path 3D Dense-UNet which took a triplet input (CBCT image, CT image, and PB?CT? contour) to yield the predicted PB?CBCT? contour corresponding to the input CBCT. The patient cohort was randomly split into 12 patients for training, 2 patients for validation, and 3 patients for testing. Each patient had 5 CBCT image-contour sets. The multi-path model was compared with a 3D UNet and a single-path model using the Dice coefficient. Friedman test for repeated measures and Wilcoxon signed-rank tests were conducted to test performance differences.
Results: The UNet, the single-path Dense-UNet, and the multi-path Dense-UNet achieved the Dice coefficient of 0.780±0.065, 0.814±0.068, and, 0.837±0.073, respectively. The p-value for the Friedman test was 0.001. P-values for all Wilcoxon signed-rank tests were less than 0.017 except the one for comparing the U-Net and single-path model.
Conclusion: Our proposed multi-path 3D Dense-UNet generated PB?CBCT? contours with the highest clinical agreement. Statistical tests indicated significant differences between the multi-path model and the 3D UNet or single-path model. This pilot study demonstrates that the proposed model is a promising tool for facilitating online contouring in the adaptive workflow. More patient data is needed to examine model robustness.
Funding Support, Disclosures, and Conflict of Interest: This study was funded by Varian Medical Systems, Inc.
Keywords
Segmentation, Cone-beam CT, Prostate Therapy
Taxonomy
IM/TH- Cone Beam CT: Machine learning, computer vision
Contact Email
