A similarity learning network for unsupervised deformable brain MRI registration

Deformable image registration is a fundamental task in medical image analysis, aiming to accurately align medical images from different time points or patients. In recent years, learning-based unsupervised deformable registration has received significant attention due to its fast end-to-end registration capability. With the development of deep learning, deformable registration networks that use various advanced network architectures also shown increasingly better registration performance. However, most recent methods have mainly focused on replacing specific layers in networks with advanced network architectures such as Transformers, without specifically addressing the key issues of feature extraction and matching in the registration task itself. In this paper, we explore the key reasons for improving registration performance using Transformers and propose a novel similarity learning network (SLNet) for unsupervised deformable brain MRI registration. In SLNet, we propose: (i) a dual-stream encoder with saliency feature enhancement (SFE) that independently extracts hierarchical features from each image using a dual-stream structure and identifies salient features by computing similarity matrices within features, and (ii) a progressive decoder with similarity feature matching (SFM) that achieves explicit feature matching by computing similarity matrices between features and progressively estimates the final deformation field in a coarse-to-fine manner. Comprehensive experiments are conducted on four publicly available 3D brain MRI datasets (OASIS, IXI, Mindboggle, and LPBA). The results demonstrate that our SLNet achieves state-of-the-art performance, with a DSC improvement of at least 4.7% and an ASD reduction of at least 0.2 mm compared to the representative VoxelMorph.