Decoupling detection and description with dual structural consistency filtering for robust multimodal retinal image registration

Registration of multimodal retinal images, such as color fundus photography (CFP) and fluorescein angiography (FA), is crucial for linking structural lesions to perfusion, guiding treatment, and monitoring disease progression. However, existing methods often struggle with large appearance gaps, sensitivity to segmentation errors, and the tight coupling of keypoint detection and description. To address these challenges, we propose a novel framework that decouples keypoint detection from descriptor extraction. Our method begins by detecting vessel bifurcations as stable anatomical landmarks directly on the original images, which prevents errors induced by segmentation. It then extracts robust descriptors from a common vessel segmentation domain, effectively bridging the modality gap. Furthermore, we introduce a Dual Structural Consistency Filter (DSCF) to prune geometrically inconsistent matches prior to homography estimation. Evaluated on two public datasets and an internal dataset, our framework achieves state-of-the-art or comparable performance across multiple standard metrics. Ablation studies show that the decoupled design enhances robustness, while DSCF improves the inlier rate and reduces the computational cost of RANSAC-based homography fitting. Our results also demonstrate that a sparse set of accurate anatomical landmarks is more effective than dense, generic keypoints. This work presents a practical and robust solution that separates anatomical landmark detection from feature description, offering a new paradigm for multimodal retinal registration.