EyeLiner: A Deep Learning Pipeline for Longitudinal Image Registration Using Fundus Landmarks

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Longitudinal fundus imaging is crucial for tracking chronic eye diseases like glaucoma and macular degeneration.
• Manual image review is subjective and challenging due to variations in image acquisition.
• Accurate image alignment is needed to distinguish true disease progression from acquisition artifacts.

Purpose of the Study:

• To introduce EyeLiner, a deep learning (DL) pipeline for registering (aligning) 2D color fundus photos (CFPs).
• To improve the alignment of longitudinal image pairs, compensating for camera variations while preserving pathological changes.

Main Methods:

• EyeLiner uses a DL-based keypoint matching algorithm to register a "moving" image to a "fixed" image.
• Anatomical keypoints on retinal blood vessels were detected using a convolutional neural network and matched with a transformer-based algorithm.
• Transformation parameters were learned from corresponding keypoints.

Main Results:

• EyeLiner demonstrated effective alignment on three datasets (FIRE, CORIS, SIGF) via qualitative and quantitative evaluations.
• Mean distance (MD) significantly decreased post-alignment: FIRE (321.32 to 3.74 px), CORIS (9.86 to 2.03 px), SIGF (25.23 to 5.94 px).
• Achieved superior Area Under Curve (AUC) scores (0.85, 0.94, 0.84) compared to state-of-the-art SuperRetina (0.76, 0.83, 0.74).

Conclusions:

• The EyeLiner pipeline offers improved image pair alignment compared to current state-of-the-art methods across multiple datasets.
• This advancement is expected to aid clinicians in aligning images and enhancing visualization of disease progression over time.