Abstract—Deformable Medical Image Registration (DMIR) is an important tool in medical imaging because it aligns images from several modalities or periods, resulting in a more accurate diagnosis and treatment planning. Single- and double-stream architectures frequently struggle with the challenges of patient organ movements and artifacts from numerous scanning devices. The study presents a transformer-based triple-stream network that uses hierarchical dual-attention techniques to enhance the relationships between moving-fixed and augmented moving-fixed images. We produced two different deformation fields: one with moving and fixed images and the other with augmented moving and fixed images using multi-scale deformation. To improve the overall registration accuracy, we used advanced motion-correction techniques to combine these fields to create the final deformation field. The effectiveness of the proposed model was validated using three benchmark datasets: OASIS, LPBA40, and Learn2Reg (Lung CT). The results show that our technique consistently outperforms learning-based methods, with outstanding results for both single and cross-modalities in medical image registration. 
 

Keywords—deformable image registration, transformer, multi-scale deformation, attention mechanism, unsupervised learning