A Retinal Image Sequence Registration Method Based on Longitudinal 3D Fundoscopy Scene Modeling

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Accurate global registration of retinal images is crucial for longitudinal studies of eye diseases.
• Existing methods often fail due to not accounting for inherent spatiotemporal transformations, leading to low accuracy and reliability.

Purpose of the Study:

• To propose a novel 3D spatiotemporal model to simulate fundoscopy scenes on a longitudinal time-axis.
• To develop an improved retinal image registration framework addressing limitations of current approaches.

Main Methods:

• A dynamic eyeball model simulating eyeball shape changes (e.g., myopia) and a camera array model simulating changes in eyeball pose and camera parameters were developed.
• A joint frame-to-reference and frame-to-frame registration strategy using matched keypoints, pose estimation, and optimization was implemented.
• The framework generates realistic synthetic retinal image data from a single image by leveraging the spatiotemporal simulation.

Main Results:

• The proposed method achieved state-of-the-art registration accuracy and reliability in comprehensive experiments on multiple datasets.
• The framework demonstrated superior performance compared to existing methods for both image pair and sequence registration.
• The ability to generate synthetic data enhances the applicability and robustness of the registration method.

Conclusions:

• The developed 3D spatiotemporal model and registration framework significantly advance the accuracy and reliability of longitudinal retinal image analysis.
• This approach provides a robust solution for fundus image registration, crucial for monitoring eye diseases over time.
• Publicly releasing codes and datasets will foster further research in retinal image analysis and ophthalmology.