Generalizable Multimodal Retinal Image Registration via Label-free Vessel Segmentation

Multimodal retinal imaging plays a crucial role in diagnosing and managing various retinal diseases such as diabetic retinopathy and age-related macular degeneration (AMD). The majority of retinal imaging modalities including Color Fundus Photography (CF), Fluorescein Angiography (FAG), Fundus Autofluorescence (FAF), Indocyanine Green Angiography (ICG), and Optical Coherence Tomography (OCT), along with infrared imaging (IR) and B-scans, capture different aspects of the same pathology within the retina. Consequently, accurate disease quantification, monitoring, and automated diagnosis require integrating their complementary insights, which relies on accurate registration of these images. To this end, developing a robust registration algorithm applicable across all modalities assumes significance. This paper presents a generalizable, label-free approach to retinal image registration, using vessel structure, extracted using DexiNed algorithm. The proposed method was evaluated across various imaging modalities, including CF-IR, CF-FAF, CF-FAG, CF-ICG, FAF-FAG, FAF-ICG, and FAG-ICG, achieving a mean landmark error (MLE) ranging from 1.91±0.44 to 4.9±2.32 pixels. In particular, CF-IR and CF-FAF registration attained an MLE of 3.08±1.47 and 4.9±2.32 pixels respectively, performing favorably when compared to human grader annotations. Furthermore, the proposed solution eliminates the need for large labeled training datasets while effectively extracting vessel structures to enable multimodal retinal image registration, improving diagnostic precision and disease monitoring.