Lightweight hybrid Mamba2 for unsupervised medical image registration

BACKGROUND: Deformable medical image registration is a critical task in medical imaging-assisted diagnosis and treatment. In recent years, medical image registration methods based on deep learning have made significant success by leveraging prior knowledge, and the registration accuracy and computational efficiency have been greatly improved. Models based on Transformers have achieved better performance than convolutional neural network methods (ConvNet) in image registration. However, their secondary computational complexity leads to significant computational overhead, posing substantial challenges for deployment in resource-constrained medical environments. Recently, Mamba-2 introduced the structured state-space Duality (SSD) framework to address the high computational cost of Transformer, achieving state-of-the-art performance across multiple domains. Mamba-2 may be a more powerful competitor than Transformer in the field of image registration. The design of its global receptive field and linear computational complexity enable it to show substantial advantages and efficiency in accurately understanding the nonlinear spatial relationships between the moving images and fixed images. PURPOSE: To address the challenges of deployment in resource-constrained medical environments and further improve the efficiency and accuracy of medical image registration, we propose HybridMorph, a lightweight hybrid Mamba2 model for medical image registration. In this study, we also introduce three versions of HybridMorph with different parameter numbers. METHODS: We propose a Residual Hybrid Module (RHM) that reconstructs a feature extraction module for medical image registration tasks based on convolution and Mamba-2, along with a novel lightweight method called the parallel channel feature aggregator (PCFA), which extracts richer feature representations with lower computational overhead. RESULTS: The proposed model was evaluated by comparing it with various existing baseline registration methods. The results show that HybridMorph achieves significant performance improvements over the baseline methods, achieving the highest average Dice scores of 0.780 and 0.824 in atlas-to-patient and inter-patient brain Magnetic Resonance Imaging (MRI) registration, respectively. Notably, compared to the renowned TransMorph, HybridMorph achieves superior registration performance while reducing the number of parameters and computational cost by 10.1 and 5.8 times, respectively. CONCLUSIONS: HybridMorph brings significant performance improvements and lower computational overhead compared to baseline methods, demonstrating the potential of our model in promoting the lightweight design of medical image registration models.