Disentangling Patient-Specific Canonical Anatomy and Deformation Manifolds for Improved MRI-Guided Adaptive Radiotherapy Segmentation
Abstract
Purpose
Accurate organ-at-risk segmentation remains a critical bottleneck in MR-guided adaptive radiotherapy, consuming 20–40 minutes per fraction. Current methods treat each fraction independently, discarding patient-specific information from prior sessions. We developed Shape-Transform-Decoder (STD), a longitudinal approach that disentangles images into patient-specific shape manifolds and deformation manifolds. We hypothesize that leveraging patient-specific prior contours to anchor canonical anatomy and suppress non-physical, spurious artifacts, while modeling daily deformation, will improve boundary accuracy and reduce contouring outliers, particularly for deformable bowel structures.
Methods
We retrospectively analyzed 104 pancreatic cancer patients (520 MRI fractions) under IRB exemption, using patient-level splits: 72 training (360 fractions), 10 validation (50 fractions), 22 testing (110 fractions). STD employs dual encoder-decoder paths: (1) Shape Encoder projects prior masks onto 256-dimensional manifolds encoding canonical topology; (2) Transform Encoder maps current images with prior masks to deformation manifolds. Manifolds fuse before the Transform Decoder with skip connections for topologically constrained predictions. Four-component loss enforces manifold orthogonality. We evaluated Dice Similarity Coefficient (DSC), 95th-percentile Hausdorff Distance (HD95), and Average Symmetric Surface Distance (ASSD) for four organs-at-risk (colon, duodenum, small bowel, stomach), comparing against five baselines: 3D U-Net, nnU-Net, SwinUNETR-V2, UNet++, and AttUNet.
Results
STD achieved mean DSC 87.31±3.93%, HD95 5.82±4.11 mm, and ASSD 1.06±0.44 mm on 22 test patients, outperforming all baselines (DSC 81.43-83.11%; p<0.001). Versus the best baseline UNet++ (83.11±7.03%), STD improved DSC by 4.2 points and reduced HD95 68% (18.18 to 5.82 mm) and ASSD 69% (3.39 to 1.06 mm). STD reduced variability 44% across cases (SD 3.93% vs 7.03%). Organ
Results
colon 90.7% DSC/4.68 mm HD95, stomach 91.6%/4.95, small bowel 85.7%/7.63, duodenum 81.3%/5.99. Mean inference time per scan was 1.03 s.
Conclusion
STD introduces the disentangled representation manifold learning framework for adaptive radiotherapy segmentation, achieving significant boundary accuracy improvement over single-fraction methods.