Rotation Equivariant Convolutions in Deformable Registration of Brain MRI

NeurIPS 2025 Workshop NeurReps Submission109 Authors

Keywords: Image registration, CNN, Steerable kernel, rotation equivariance

TL;DR: We investigate the impact of replacing standard convolutions with rotation equivariant convolutions in deformable brain MRI registration networks.

Abstract: Image registration aligns anatomical structures between pairs of medical images. Deep learning-based registration methods have achieved state-of-the-art performance using convolutional neural networks (CNNs). However, while CNNs are translation equivariant, they lack rotation equivariance. This limitation prevents networks from fully exploiting the inherent rotational symmetries present in anatomical structures, particularly in brain MRI where these symmetries are prominent features. In this work, we investigate the impact of replacing standard convolutions with rotation equivariant convolutions in deformable brain MRI registration networks. We evaluate our approach on two baseline architectures (VoxelMorph and Dual-PRNet++) across multiple brain MRI datasets and compare against a non-symmetric control dataset. Our experiments demonstrate that rotation equivariant encoders improve registration accuracy on symmetric brain data while showing decreased performance on non-symmetric anatomical structures, confirming that the inductive bias of rotational symmetry is beneficial when anatomically justified.

Submission Number: 109