(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purpose: To develop a 3D deep-learning model for the automatic segmentation of gross tumor volumes (GTVs) for palliative head and neck radiotherapy planning.
Methods: 104 palliative treatment plans (non-contrast CT scans and physician-approved GTV contours) were retrospectively curated and split into datasets of 68-18-18 train, cross-validation, and test, respectively. A 3D U-Net model was trained to autosegment the GTVs. The model performance was evaluated using Dice similarity coefficient (DSC) and mean surface distance (MSD) comparing the predicted and clinical GTVs. Because most palliative radiotherapy plans use opposed lateral fields, the model performance was also evaluated on lateral projections of the GTVs. Finally, treatment plans were created for the test patients using the same gantry angles as the original plan (typically opposed laterals), and a GTV-to-block-edge expansion of 1.5cm. The dosimetric coverage of the original clinical contours was evaluated.
Results: Of the 18 test cases, inferences failed for 2 (DSC<0.2). For the remaining 16 cases, the mean 3D-DSC comparing the predicted and clinical GTVs was 0.66 (range 0.44-0.87, s 0.12); mean MSD was 4.6mm (range 1.9-8.8mm, s 1.8). The mean 2D-DSC for the lateral projections of the GTVs was 0.77 (range 0.56-0.93, s 0.10). When treatment plans were prepared using the predicted GTV, the clinical GTV contours received a minimum 95% prescription dose for 14/16 of these cases (88%).
Conclusion: While challenging, a deep-learning model for auto-segmentation of GTVs has potential for automatic planning of palliative head and neck radiotherapy treatments.
Funding Support, Disclosures, and Conflict of Interest: Disclosures: Our research group receives funding from the NCI and Varian Medical Systems.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
