Author: Christopher G. Ainsley, Pradeep Bhetwal, Yingxuan Chen, Wookjin Choi, Vimal K. Desai, Karen E. Mooney, Adam Mueller, Hamidreza Nourzadeh, Yevgeniy Vinogradskiy, Maria Werner-Wasik 👨🔬
Affiliation: Thomas Jefferson University 🌍
Purpose: MR-guided adaptive radiotherapy (MRgART) has demonstrated improved outcomes for patients with pancreatic cancer. However, the time-consuming re-segmentation of targets and organs-at-risk (OARs) for each MRgART treatment prolongs the overall treatment time. This study evaluated state-of-the-art deep learning (DL) segmentation models for automatically segmenting targets and OARs for MRgART.
Methods: We collected a total of 126 image sets, including planning MR and daily MR scans, from 21 pancreatic cancer patients undergoing MRgART. Three deep learning (DL) model architectures, SegResNet, SegResNet 2D, and SwinUNETR were used in this study. The models automatically segment targets (GTV, CTV, and PTV) and OARs (Small and Large Bowels, Duodenum, Left and Right Kidneys, Liver, Spinal Cord, and Stomach) if they lie within the adaptive ring. The model was trained on a set of planning MR and fraction 1-4 MR images and evaluated on fraction 5 segmented MR images using 5-fold cross-validation. Six metrics were used to evaluate model performance: average surface distance (ASD), Hausdorff-95, Hausdorff-100, surface overlap, surface Dice (SDice) and volumetric Dice (VDice).
Results: SwinUNETR demonstrated superior performance across both datasets, achieving the lowest surface distances and the highest overlap and Dice scores. SwinUNETR achieved the average VDice (0.73±0.14) and average SDice (0.43±0.15), GTV (VDice: 0.73±0.19, SDice: 0.40±0.18), the best structure: Kidney (VDice:0.92±0.04, SDice:0.61±0.12) and the worst: Small Bowel (VDice: 0.63±0.15 , SDice: 0.22±0.10) in the testing dataset. SegResNet 2D consistently showed higher ASD (5.70±7.44) and lower accuracy across both datasets.
Conclusion: SwinUNETR consistently outperformed both SegResNet and SegResNet 2D on all evaluation metrics. SwinUNETR could be a valuable tool for improving the efficiency and accuracy of re-delineation in MRgART for pancreatic cancer patients, demonstrating its potential to reduce a significant workflow bottleneck and improve efficiency for patients.