Architectural and Regularization Components in Deep Learning Medical Image Registration: Systematic Ablation Study

Systematic ablation study quantifying individual and synergistic effects of architectural enhancements versus regularization losses in deep learning-based medical image registration. Using the OASIS brain MRI dataset (n=414; 394 training, 20 test), we demonstrate that regularization losses are the primary performance driver, achieving 21.3% improvement in registration accuracy with 99% reduction in unrealistic deformations and negligible computational overhead. Combined with affine architectural components, the approach achieves 25.8% total improvement with sub-voxel accuracy and anatomically plausible deformation constraints suitable for clinical deployment.