Study Isolates Key Drivers in Deep Learning Medical Image Registration

A controlled ablation study released on arXiv has evaluated the distinct contributions of architectural enhancements and regularization losses in deep learning-based medical image registration. Using the OASIS brain MRI dataset comprising 394 training and 20 test subjects, the research examined four variants of a 3D U-Net registration pipeline.

The findings show that regularization alone accounted for most performance improvements, delivering a 21.3% relative gain in MSE improvement and a 21.8% gain in NCC. Crucially, regularization reduced unrealistic maximum deformation by 99.0% (from 53.1 to 0.51 units) with virtually no inference-time computational cost (-0.06%).

The fully enhanced model, which combines an affine pre-alignment architecture with complete regularization, achieved the highest overall accuracy gain of 25.8% and increased anatomical plausibility from 0.596 to 0.930 at a modest +9.8% inference time cost. The authors emphasize that eliminating unrealistic deformations directly addresses a primary barrier to clinical deployment.