Purpose: Fast and consistent GTV segmentation in non-contrast enhanced daily 4DMRI is challenging for MR-guided Online Adaptive Radiation Therapy (MRgOART) for abdominal tumors. In this work, we develop a prior knowledge guided deep learning model specifically for auto-segmentation of pancreatic GTV on daily 4DMRI for MRgOART.
Methods: Seventy-three pairs of verification and daily mid-position MRI sets acquired from 12 pancreatic cancer patients using 4D Vane sequence on a 1.5T MR-Linac during MRgOART were used to train and test the deep learning model with patient-based leave-one-out cross-validation. The model used the verification MRI and GTV mask (obtained before the first fraction) as a guidance for the daily MRI. A region of interest (ROI) was selected as the minimum bounding box of the verification GTV contour plus a margin of 15 mm to account for inter-fractional variations. Model input size was 128×128×3 with 3 channels corresponding to verification MRI, verification GTV mask, and daily MRI. Data augmentation methods such as translation, rotation, flip, guidance-target inversion, and rigid registration error introduction were applied for robust model training. Model performance was evaluated using dice similarity coefficient (DSC).
Results: With the small size of the 4D Vane MRI dataset, the proposed model achieved an average DSC of 0.66±0.13 for all the cases studied. While for the 8 cases with GTV >5 cc, the average DSC was 0.71, comparable to the inter-observer variation of DSC = 0.71 estimated in a previous study. The proposed model took less than 1 second to create the GTV contour on a daily MRI set.
Conclusion: We developed a prior knowledge guided deep learning model to segment the pancreatic GTV in the daily non-contrast enhanced MRI using a composite image-mask input. The proposed model achieved clinically relevant DSC for cases with GTV >5 cc.
Not Applicable / None Entered.