Coarse-to-fine medical image registration with landmarks and deformable networks

Medical image registration is a critical task in medical image processing. However, most deep learning-based registration methods primarily rely on a single network to directly predict the deformation field, which can result in challenges when accurately matching certain complex imaging regions. Moreover, the pervasive issue of insufficient training data in medical image registration further limits improvements in model performance. To address these challenges, we propose a novel coarse-to-fine image registration architecture (CFIR) consisting of a landmark-based registration network (LRN) and a deformable registration network (DRN). LRN first extracts and aligns contours and key anatomical regions from the images, achieving coarse registration. Subsequently, DRN refines the alignment by concentrating on finer details that were not adequately addressed during the coarse registration stage, ensuring precise alignment of local features between the images to be registered. During the training process, we employ segmentation and recombination to augment the training data. The performance of CFIR is rigorously evaluated on brain MRI and abdominal CT registration tasks, demonstrating superior registration accuracy compared to several existing CNN-based and Transformer-based methods. Our method provides a new paradigm for medical image registration.