MFCTrans: Multi-scale Feature Connection Transformer for Deformable Medical Image Registration

Abstract

Deformable Medical Image Registration (DMIR) aims to establish precise anatomical alignment of multiple medical images. However, the existing U-shape networks encounter difficulties in efficiently transferring multi-scale feature information from the encoder to the decoder. To address this issue, we propose a novel backbone network called MFCTrans, which constructs effective feature connection in DMIR. Drawing inspiration from the attention mechanism observed in the human cognitive system, our proposed method employs a Feature Fusion and Assignment Transformer (FFAT) module and a Spatial Cross Attention Fusion (SCAF) module. The former facilitates the fusion of multi-channel features, while the latter guides the integration of multi-scale information. A Multiple Residual (MR) branch is also deployed between the encoder and FFAT to improve the network’s generalization. We conduct extensive qualitative and quantitative evaluations on the OASIS and LPBA40 datasets. The proposed method achieves higher Dice scores than Transmorph by 1.3% and 2.0% on the respective datasets while maintaining a comparable voxel folding percentage. Ablation studies analyze the impacts and efficiency of each component in the proposed method. In summary, our proposed network offers a promising framework for achieving high-quality medical image registration and holds significant potential for applications in computer vision and cognitive computation.