DMSC-Reg: Medical Image Deformable Registration Based on Dynamic Multi-Scale Convolutional Fusion

This study explores the application of deep learning techniques in medical image registration and introduces a novel hybrid CNN-Transformer architecture that incorporates dynamic convolutional fusion. The proposed approach harnesses dynamic multi-scale convolution to overcome the inherent limitations of the receptive field in traditional convolutional neural networks while simultaneously addressing the computational complexity and parameter redundancy that often constrain the practical deployment of Transformer-based registration models. By integrating the strengths of both convolutional and Transformer architectures, our method achieves a balanced trade-off between local feature extraction and global context modeling. Experimental evaluations on the publicly available IXI dataset demonstrate that our model achieves superior performance, with a significantly higher Dice coefficient score compared to state-of-the-art deep learning registration algorithms. These results underscore the effectiveness of our approach in enhancing the precision and robustness of medical image registration.