Distance-to-Contour Conditioned Coregistration for Pancreatic Cancer RT
Abstract
Purpose
Longitudinal deformable registration is crucial for accurate dose accumulation in magnetic resonance guided adaptive radiotherapy (MRgART). This is particularly difficult for pancreatic cancer, with both substantial inter-fraction spatial shifts and intensity variation from food and gas. This motivates creation of deformable methods robust to intensity shifts.
Methods
We propose a distance-to-contour conditioned coregistration (DisCo3) method based on ProRSeg. ProRSeg iteratively refines organ segmentations using the registration. With DisCo3 we concurrently inform the registration with contour conditioning, after converting organ masks to surface distance maps. This model segments target fraction organs-at-risk (OAR) while deformably registering prior fractions to the target fraction, with each task aiding the other. Training consisted of 30 patients with 5 scans corresponding to 5 fractions of RT (150 scans total). The model was evaluated on 9 withheld patients. DSC, HD95, SSIM, Jacobian determinant, and bending energy were computed for every pairing of fractions (180 registrations) and significance was evaluated using a Wilcoxon signed-rank test.
Results
Compared to ProRSeg, DisCo3 significantly improved DSC and HD95 alignment. ProRSeg DSC was 0.95, 0.75, 0.62, and 0.80 for liver, large bowel, small bowel, and stomach-duodenum, while DSC for DisCo3 was 0.96 (p<0.001), 0.87 (p<0.001), 0.72 (p<0.001), and 0.84 (p=0.034), respectively. ProRSeg HD95 was 4.0, 13.7, 15.1, and 6.4 mm, for liver, large bowel, small bowel, and stomach-duodenum, while HD95 for DisCo3 was 3.1 (p<0.001), 6.0 (p<0.001), 11.2 (p<0.001), and 6.2 mm (p=0.02), respectively.
Conclusion
While ProRSeg demonstrated the effectiveness of improving segmentation through iterative refinement with joint registration, we show with DisCo3 that this works symmetrically, with progressively refined segmentation information enhancing registration – particularly when providing this segmentation in the richer feature space of a surface distance map. Use of a surface distance map also enables registration using manually edited contours in place of the segmentation module.