A coarse-to-fine registration method for multimodal retinal images

Manual preoperative image registration for central serous chorioretinopathy (CSCR) is labor-intensive and irreproducible. While rigid registration robustly aligns images globally, it misses fine details. Non-rigid registration, though excellent for local refinement, performs poorly with large discrepancies. Therefore, this study presents a coarse-to-fine registration method for multimodal retinal images to address the aforementioned issues. First, a three-step coarse registration strategy is designed that integrates keypoint pair detection and matching via a YOLOv8-pose network, further optimizes keypoints through a post-processing technique, and achieves initial alignment via affine transformation. On this basis, a dual-component fine registration strategy is then implemented, where disentanglement learning eliminates modality-specific variations while preserving essential vessel structures required for registration, and deformable network generates optimized deformation field to refine the coarse alignment locally, ultimately enabling high-precision image registration. Comprehensive qualitative and quantitative experiments were conducted on the CSCR clinical dataset, which includes both color fundus (CF) and fundus fluorescence angiography (FFA) images, to evaluate the proposed method. With Dice and Dice_s scores of 0.6759 and 0.4977, the method performs comparably to existing approaches, suggesting its potential application value for CSCR preoperative planning.