Segmentation regularized training for multi-domain deep learning registration applied to magnetic resonance-guided prostate cancer radiotherapy

Background and purpose: Accurate deformable image registration (DIR) is required for contour propagation and dose accumulation for magnetic resonance guided adaptive radiotherapy (MRgART). The goal of this study was to train segmentation regularized deep learning (DL) DIR and assess domain invariant MR-MR registration. Materials and methods: DL-DIR based progressively refined registration and segmentation (ProRSeg) and VoxelMorph were trained with and without organ weighted segmentation regularization with 262 MR pairs from same domain dataset of longitudinal 3 Tesla MR simulation (MR-Sim) scans of patients with prostate cancer acquired before and every 3 months following radiotherapy. Models were compared against variational DIR by measuring accuracy of organs and clinical target volume (CTV) contour propagation accuracy on same- (58 pairs), cross- (72 1.5 Tesla MR-Linac pairs from consecutive daily treatment fractions), and mixed-domain (42 MRSim planning − MR-Linac first fraction pairs) datasets. Dose accumulation was performed for 42 patients undergoing 5-fraction MRgART on MR-Linac. Results: ProRSeg produced robust DIR on cross- (all organs) and mixed-domain (bladder, CTV) datasets. It was the most accurate method for highly deforming organs including bladder and rectum. Segmentation regularization enhanced accuracy of both ProRSeg and VoxelMorph compared to unregularized versions. Dose accumulation feasibility study with ProRSeg indicated that 83.3% of patients met key institutional constraints for CTV coverage and bladder sparing. Conclusions: ProRSeg was a feasible approach for multi-domain MR-MR registration for prostate cancer patients. Dose accumulation analysis indicated preliminary feasibility to evaluate compliance of delivered treatments to clinical constraints.