Room: AAPM ePoster Library
Purpose: post-operative prostate bed for adjuvant and early salvage radiotherapy is a difficult task. Not only has normal anatomy been altered by the surgical removal of the prostate, there is a lack of visible target which can be defined. Segmentation is often based on guidelines and other clinical information. We propose a deep learning solution for segmenting CTV by segmenting nearby organs first and then using these to assist CTV segmentation. Model is trained using labels clinically approved and used for treatment, which however are subject to high inter/intra observer variability due to the absence of visual truth. We therefore quantify the associated model/data uncertainties to help physicians revise the segmented CTV.
Methods: adjuvant/salvage therapy patients were included. The proposed workflow mapped patient CT to volume-wise labels of bladder, rectum, femoral heads, penile bulb and CTV. OAR segmentations were performed first using appropriate deeplearning architectures. Since post-op CTV segmentation relies on anatomical location, a Multi-task network was used with distance prediction as an auxiliary task. Uncertainty was estimated using MonteCarlo Dropout(MCDO). The dice similarity coefficient (DSC) was calculated between the physician-contoured CTV and model-generated CTV. Physicians then scored both the physician- and model-generated CTVs on a 4-point grading system.
Results: achieved a DSC of 0.86 for CTVs and state-of-the-art result for OARs on the test dataset (50 patients). Equivalence test for means between the model- and clinical-contour scores from physician evaluation showed that they are statistically equivalent (p<.00001). Uncertainties estimated were highly correlated to the error in prediction for OARs. For CTV, they provide valuable information regarding inter/intra-observer variability.
Conclusion: have proposed a deeplearning model to delineate CTV and OARs on post-operative prostate CT which demonstrated high levels of similarity and were indistinguishable from the physician-generated CTVs. For real-time interpretation of the contours, MCDO was found to be effective.
Funding Support, Disclosures, and Conflict of Interest: Funding Support: VARIAN Research Grant