JMIR Publications

ABSTRACT

Deep learning-based medical image registration methods increasingly incorporate both architectural enhancements (affine transformations) and training objective improvements (regularization losses), yet their individual and combined contributions remain poorly understood. To quantify the individual and synergistic effects of affine components versus regularization losses on deformable medical image registration performance through systematic ablation analysis, we conducted a controlled ablation study using the OASIS brain MRI dataset comparing four model variants: baseline 3D U-Net with basic similarity losses, regularization-enhanced U-Net, affine-enhanced U-Net with basic losses, and fully enhanced model combining both components. Primary outcomes included registration accuracy metrics (mean squared error [MSE], normalized cross-correlation [NCC], structural similarity index [SSIM]), enhanced deformation quality analysis including Jacobian determinant preservation and anatomical plausibility scoring, and computational efficiency measures. Regularization enhancement alone achieved substantial performance improvements: 21.3% relative improvement in MSE (1.78% → 2.16%, P<.05) and 21.8% improvement in NCC (0.0555 → 0.0676), while dramatically reducing maximum deformation from 53.1 to 0.51 units (99.0% reduction) with negligible computational overhead (-0.06% inference time). Combined approaches achieved optimal performance with 25.8% relative MSE improvement (1.78% → 2.24%) and enhanced anatomical plausibility scores (0.596 → 0.930), at moderate computational cost (+9.8% inference time). Enhanced gradient correlation analysis revealed substantial improvements in structural preservation (0.742 → 0.980 for fully enhanced model). All enhanced variants achieved sub-voxel registration accuracy with anatomically plausible deformation constraints. Regularization losses provide the primary driver of performance improvements in medical image registration, offering both accuracy gains and dramatic deformation control enhancement with maintained computational efficiency. Architectural enhancements provide complementary benefits at acceptable computational cost. The dramatic improvement in deformation control (99% reduction in unrealistic deformations) addresses critical clinical deployment concerns while achieving superior registration accuracy.