Evaluation of classical and deep-learning deformable registration for magnetic resonance to four-dimensional computed tomography contour mapping in liver stereotactic body radiotherapy.

Background and purpose: : In liver stereotactic body radiotherapy (SBRT), magnetic resonance imaging (MRI)-to-four-dimensional computed tomography (4D-CT) contour mapping remains challenging because of cross-modality differences and respiratory motion, motivating fast automated solutions. We comparatively evaluated the clinical practicality of deep learning (DL) based deformable image registration (DIR). Materials and Methods: : This retrospective study included 4D-CT and 3D MRI scans from 170 patients treated with liver SBRT. Segmentations were obtained using open-source automated tools, and test-set labels were clinician reviewed. The dataset was split at the patient level into training/validation/testing sets (122/32/16), an additional 5-fold cross-validation was performed on the training+validation cohort. We compared three approaches: an optimization-based B-spline DIR tool (NiftyReg), a single-network DL model (Direct), and a two-step DL pipeline decomposing multi-modal alignment and intra-4D-CT motion tracking. Statistical analysis used patient-level paired Wilcoxon signed-rank tests; organ-wise secondary analyses were corrected using the Benjamini-Hochberg false discovery rate (FDR; q=0.05). Results: Classical B-spline DIR achieved the best all-label averages (Dice 0.71 vs 0.62; 95th percentile Hausdorff distance [H95], 17.7 mm vs 21.0 mm). DL achieved its strongest performance on supervised structures compared to all labels (e.g., Dice 0.71 vs 0.62; H95 18.9 mm vs 21.0 mm) while reducing runtime from Conclusion: DL pipelines provided competitive contour-mapping accuracy with orders-of-magnitude runtime gains, but performance depended strongly on the choice of training supervision; classical DIR remained strongest on global all-label metrics.