MATGAN: a generative adversarial network with multi-scale attention and Huber loss for brain MRI registration

Brain magnetic resonance image registration is a fundamental task in medical image analysis. However, achieving fine-grained alignment remains challenging, particularly in brain regions with small volumes or complex anatomical structures. Generative adversarial networks (GANs) have provided new opportunities to improve registration accuracy through adversarial learning. In this study, we propose MATGAN, a novel GAN-based registration framework that integrates a multi-scale attention mechanism and Huber loss into the adversarial learning process to achieve more robust and accurate image alignment. Specifically, a multi-scale attention module is embedded into the discriminator to extract semantic representations at different resolutions, thereby enhancing its ability to distinguish real deformations from generated ones. The similarity feedback provided by the discriminator is further transferred to the registration network, guiding it to generate deformation fields with improved anatomical consistency. In addition, Huber loss is incorporated into the similarity constraint to reduce the influence of outliers during optimization and improve the robustness and stability of network training. Through the adversarial interaction between the registration network and the discriminator, MATGAN enables more precise deformation estimation and better structural alignment. Experimental results on three public brain MRI datasets demonstrate that MATGAN achieves effective registration performance.