Retinal Image Registration and Mosaicking System for Comprehensive Wide-Field Visualization

Abstract

[Abstract]: Retinal imaging plays a crucial role in the diagnosis and monitoring of many ocular and systemic diseases, offering a non-invasive approach to analyze the retinal surface. Despite its wide range of applications, the limited field-of-view in retinal images presents a major challenge for the comprehensive assessment of the retinal surface. To address this limitation, we propose a novel retinal image registration and mosaicking system that enables the creation of seamless wide-field retinal views. The system consists of a deep learning feature-based registration framework that introduces a novel two-stage RANSAC-based outlier rejection strategy together with the use of an Approximate Thin-Plate Spline (ATPS) interpolation for robust deformable transformation estimation in retinal imaging. Our outlier rejection strategy effectively captures global and local deformations, reducing the dominance of densely matched regions which may not fully characterize the entire image, resulting in a more reliable and evenly distributed inlier set. The final deformable transformation is achieved by applying an ATPS interpolation model, which accounts for potential keypoint localization errors, ensuring a precise and flexible alignment. Finally, our blending method stitches together the images aligned to a common reference, creating smooth and seamless retinal mosaics. Evaluation on the reference FIRE dataset demonstrated that our retinal registration framework outperformed all previous state-of-the-art methods. We additionally validated its robustness across other datasets and illustrated the application of our blending method to retinal mosaicking. The results demonstrated the effectiveness of our system, and its consistency generating seamless wide-field retinal visualizations, ultimately enhancing diagnostic accuracy and treatment planning.