Leveraging Prior Fractions to Improve Segmentation in Fractionated Adaptive Radiotherapy
Recommended Citation
Li C, Rusu D, Dolan J, Parikh PJ, Thind K. Leveraging Prior Fractions to Improve Segmentation in Fractionated Adaptive Radiotherapy. Med Phys 2025; 52(8):73.
Document Type
Conference Proceeding
Publication Date
8-14-2025
Publication Title
Med Phys
Abstract
Purpose: To develop and evaluate a novel deep learning framework that leverages prior fraction information to improve organ-at-risk segmentation accuracy in fractionated adaptive radiotherapy. Methods: Data from 74 pancreatic cancer patients treated with 5-fraction magnetic resonance guided adaptive radiotherapy were analyzed. Images were pre-processed with resampling (1.5×1.5×3.0mm voxel spacing), density normalization, and cropping (128×128×64 dimensions). The proposed dual-path architecture incorporates both current fraction imaging (infer path) and prior fraction imaging with segmentation data (support path), with feature fusion blocks integrating multi-scale spatial and temporal features. Performance was evaluated using both 3D UNet and SwinUNETR backbones, comparing baseline implementations (without prior fraction data) against our sequential framework. Dice Similarity Coefficient (DSC), 95th percentile Hausdorff Distance (HD95), and Average Symmetric Surface Distance (ASSD) were assessed with Mann-Whitney U-test for statistical significance. Results: The sequential approach yielded consistent improvements across all evaluated organs (colon, duodenum, small bowel, stomach) with both network backbones. With 3D UNet, significant increases in DSC were observed for colon (0.851 to 0.871, p=0.0525) and small bowel (0.778 to 0.820, p=0.0399), while HD95 for duodenum significantly decreased (32.46mm to 26.37mm, p=0.0087). Using SwinUNETR, significant improvements were seen in DSC for duodenum (0.690 to 0.730, p=0.0129) and HD95 for stomach (26.22mm to 14.82mm, p=0.0350). Conclusions: The proposed framework effectively leverages temporal information across treatment fractions to enhance segmentation accuracy in adaptive radiotherapy, providing consistent improvements with both convolutional and transformer-based architectures, suggesting potential for clinical implementation and improved efficiency in ART workflows.
Volume
52
Issue
8
First Page
73
