TY  - JOUR
TI  - A comparative study of deep learning-based retinal image registration methods
AU  - Dharmaseelan, Thenuka
AU  - Sinha, Neelabh
AU  - Ashraf, Samyyia
AU  - Daneshvar, Kimia
AU  - John, Amit
AU  - Giannakis, Periklis
AU  - Chan, Yik Ting
AU  - Chan, Yiu Wai
AU  - Pontikos, Nikolas
PY  - 2026
JO  - Exploration of Digital Health Technologies
VL  - 4
SP  - 101194
DO  - 10.37349/edht.2026.101194
UR  - https://www.explorationpub.com/Journals/edht/Article/101194
AB  - Aim: To benchmark three deep learning-based retinal image registration methods RetinaRegNet, EyeLiner, and GeoFormer on the Fundus Image Registration (FIRE) dataset to compare registration accuracy and computational efficiency using mean landmark error (MLE) as the primary outcome measure. Methods: The three image registration approaches were evaluated using the FIRE dataset under consistent conditions across varying image overlap conditions (Classes S, A, and P). These included: (a) RetinaRegNet, which incorporates diffusion features, dual keypoint sampling through Scale-Invariant Feature Transform (SIFT) and random, two-stage outlier removal, and a multilevel registration hierarchy progressing from homography to polynomial transforms; (b) EyeLiner, which integrates anatomical segmentation with SuperPoint feature extraction, LightGlue matching, and thin-plate spline warping; (c) GeoFormer, which builds on Local Feature Transformers (LoFTR) through cross-attention mechanisms and Random Sampling Consensus (RANSAC)-based refinement. Registration performance was quantified using MLE. Results: Across all 134 FIRE image pairs, RetinaRegNet achieved the lowest overall MLE (3.12 pixels), outperforming EyeLiner (3.81 pixels) and GeoFormer (6.06 pixels). Class-specific analysis showed that RetinaRegNet delivered the highest accuracy in Class S images (1.70 pixels), competitive performance in Class A (5.24 pixels), and the strongest results in the most challenging Class P cases (4.57 pixels). GeoFormer demonstrated the shortest processing time at 0.32 seconds per image pair, compared with 4.92 seconds for EyeLiner and 31.23 seconds for RetinaRegNet. In Class P, RetinaRegNet achieved a 59.2% improvement in accuracy relative to GeoFormer (4.57 vs 11.20 pixels). The code is available at: https://github.com/ThenukaDharmaseelan/image_Registration. Conclusions: Overall, the evaluation reveals a clear trade-off between registration precision and computational speed. RetinaRegNet achieves the lowest MLE for complex clinical cases despite higher computational cost. EyeLiner balances precision and speed for routine use, while GeoFormer prioritizes rapid throughput where processing speed is critical.
ER  -