Description
Aim: To benchmark three deep learning-based retinal image registration methods RetinaRegNet, EyeLiner, and GeoFormer on the Fundus Image Registration (FIRE) dataset, comparing accuracy and computational efficiency using Mean Landmark Error (MLE) as the primary metric.
Methods: The methods were evaluated under consistent conditions across three overlap-based categories: Class S (71 pairs, >75% overlap without anatomical changes), Class A (14 pairs, >75% overlap with anatomical variations), and Class P (49 pairs, <75% overlap without anatomical changes). RetinaRegNet integrates diffusion features, dual keypoint sampling (SIFT and random), two-stage outlier removal, and a hierarchical registration strategy from homography to polynomial transforms. EyeLiner combines anatomical segmentation with SuperPoint feature extraction, LightGlue matching, and thin-plate spline warping. GeoFormer extends Local Feature Transformers (LoFTR) using cross-attention mechanisms and RANSAC-based refinement. Performance was assessed using MLE.
Results: Across 134 FIRE image pairs, RetinaRegNet achieved the best MLE (3.12 pixels), outperforming EyeLiner (3.81 pixels) and GeoFormer (6.06 pixels).Class-wise, RetinaRegNet performed best in Class S (1.70 pixels) and Class P (4.57 pixels), while EyeLiner showed comparable performance in Class A (4.87 vs 5.24 pixels). GeoFormer exhibited substantially higher errors in Class P (11.20 pixels).In terms of runtime, GeoFormer was fastest (0.32 s), followed by EyeLiner (4.92 s) and RetinaRegNet (31.23 s).
Conclusions: Results indicate a trade-off between accuracy and efficiency: RetinaRegNet provides highest precision, EyeLiner balances both, and GeoFormer prioritizes speed.
Lay Abstract
This study compared three artificial intelligence (AI) methods used to align retinal images: RetinaRegNet, EyeLiner, and GeoFormer. Retinal image alignment helps doctors compare scans taken at different times or from different devices, useful for monitoring eye diseases and treatment progress.
The three methods were tested using the FIRE dataset, a standard collection of 134 retinal image pairs with different difficulty levels. Performance was measured by how accurately each method matched key points and how quickly each method processed images.
The results showed that RetinaRegNet was the most accurate overall, achieving the lowest average alignment error. It performed especially well in difficult image pairs, where image overlap was smaller or conditions were more demanding. EyeLiner also produced strong results and offered a better balance between accuracy and speed. GeoFormer was the fastest method, making it useful for situations where many images need to be analysed quickly, although its accuracy was lower than the other two methods.
Overall, the study shows that choosing the best method depends on clinical needs. RetinaRegNet is best when high precision is required, EyeLiner is suitable for routine use, and GeoFormer is useful for high-volume rapid screening tasks.
| Lay Title | Evaluating AI Tools for Retinal Image Alignment |
|---|---|
| Role | Research Assistant |