RetinaRegNet: A zero-shot approach for retinal image registration

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Retinal image registration is crucial for monitoring eye diseases and guiding treatments.
• Existing methods face challenges with large deformations, limited overlap, and diverse image quality.

Purpose of the Study:

• To develop a robust and generalizable retinal image registration model.
• To achieve accurate alignment across multiple retinal imaging modalities with zero-shot capability.

Main Methods:

• Proposed RetinaRegNet, a multi-stage model leveraging a pretrained latent diffusion model for feature extraction.
• Employed SIFT and random sampling for feature point extraction, cosine similarity for correspondence estimation, and inverse consistency for outlier detection.
• Utilized a two-stage registration: homography for global alignment and third-order polynomial for local deformations.

Main Results:

• RetinaRegNet demonstrated superior performance across color fundus, fluorescein angiography, and laser speckle flowgraphy modalities.
• Achieved high AUC scores: 0.901 (color fundus), 0.868 (fluorescein angiography), and 0.861 (laser speckle flowgraphy).
• Showcased significant zero-shot generalizability, outperforming state-of-the-art methods.

Conclusions:

• RetinaRegNet offers a robust solution for retinal image registration, addressing key challenges in the field.
• Its zero-shot generalizability makes it a valuable tool for tracking disease progression and treatment efficacy.
• The model's performance across diverse imaging modalities highlights its clinical potential.