Author: Jennifer L. Dolan, Chengyin Li, Parag Parikh, Doris N. Rusu, Kundan S Thind 👨🔬
Affiliation: Henry Ford Health, Cedars-Sinai Medical Center 🌍
Purpose: The time and resource demands of online Adaptive Radiation Therapy (ART) can limit its widespread clinical adoption and potentially impact patient throughput. To address this, we developed a 3D UNet-based sequential segmentation framework that leverages information from prior fraction images and segmentation to enhance accuracy and efficiency of the segmentation process.
Methods: This retrospective study analyzed a dataset of 23 pancreatic cancer patients undergoing 5-fraction online adaptive MR-guided radiotherapy (MRgRT), focusing on the segmentation of four abdominal organs: colon, duodenum, small bowel, and stomach. Preprocessing included resampling to voxel spacing of 1.5 × 1.5 × 3.0 mm and cropping input volumes to 128 × 128 × 64. The proposed sequential segmentation framework integrates current and prior fraction information using a dual-path architecture with feature fusion blocks. We conducted three-fold cross-validation for model evaluation, ensuring patient-wise separation to prevent data leakage. Model performance was compared to a baseline 3D UNet without sequential support using three metrics: Dice Similarity Coefficient (DSC) for volumetric overlap, 95th percentile Hausdorff Distance (HD95) for surface distance, and Average Symmetric Surface Distance (ASSD) for boundary differences.
Results: The sequential method consistently outperformed the baseline across all organs. For the colon, DSC improved from 0.834 to 0.858 (p=0.049), HD95 decreased from 10.82mm to 8.45mm (p=0.048), and ASSD reduced from 1.73mm to 1.39mm (p=0.036). Improvements were also observed in the duodenum (DSC: 0.713 to 0.719), small bowel (DSC: 0.766 to 0.780), and stomach (DSC: 0.876 to 0.885), although not statistically significant. The most substantial improvements were consistently found in colon segmentation.
Conclusion: The proposed 3D UNet-based sequential segmentation framework effectively leverages prior fraction information to enhance segmentation accuracy and precision in adaptive radiotherapy. This approach, adaptable to other UNet-like architectures, holds significant promise for improving the efficiency and effectiveness of clinical segmentation tasks, ultimately benefiting patient care.