Deformable registration network based on multi-scale features and cumulative optimization for medical image alignment

When faced with significant structural differences or topological changes in medical image registration, existing deep learning methods still encounter challenges. To address the issue of large deformations in medical image registration, this paper proposes a deformable registration network based on multi-scale features and cumulative optimization. A multi-scale feature fusion module is designed in the encoding phase of the network to extract features from different dimensions and merge them into a high-dimensional representation. Simultaneously, an attention mechanism is employed to enhance registration accuracy in local anatomical regions. Additionally, a sustainable optimization mechanism for deformation field prediction is introduced, incorporating long-range and short-range information within the network to generate the final spatial displacement vectors. This strengthens the model's ability to handle large deformations. This framework is proven through extensive experiments to be an efficient medical image registration scheme.