Retinal image registration is crucial for tracking eye diseases and planning treatments. However, it's tricky due to significant changes in the images, limited overlap, and varying quality. Enter RetinaRegNet, a model that can handle different types of retinal images without needing to be retrained each time. It starts by using a pretrained latent diffusion model to find important features in the images. It then picks points from one image and finds matching points in the other using cosine similarities. To make sure the matches are accurate, it checks for consistency and removes any that don't fit. The remaining points help calculate how one image should be transformed to match the other. RetinaRegNet does this in two stages: first, it aligns the images globally using a homography, and then it adjusts for local changes with a third-order polynomial transformation. Tested on three types of images—color fundus, fluorescein angiography, and laser speckle flowgraphy—RetinaRegNet outperformed other methods, scoring high in all. This zero-shot capability makes it a promising tool for monitoring disease progression and treatment effectiveness. The code is available for anyone to use.

Easy Image Matching: A New Way to Track Eye Diseases

Actions