Retinal cross-modal registration skips segmentation: MICRO benchmark and one-step OFS2Reg reach 0.784 success rate

Cross-modal retinal image rigid registration now has a public benchmark covering all ophthalmic surgery phases and a one-step registration route that bypasses the vessel segmentation stage. Earlier retinal registration datasets were mostly limited to a single modality or a fixed dual-modality pairing and lacked cross-modal pairs spanning preoperative, intraoperative and postoperative phases; mainstream deep registration methods followed a segmentation-then-detection two-step pipeline. MICRO holds 1,365 cross-modal image pairs split into a preoperative subset of 1,108 pairs from 305 eyes, a pre-to-intraoperative subset of 207 pairs from 10 eyes, and a pre-to-postoperative subset of 50 pairs from 50 eyes, with 10 corresponding control points per pair annotated by expert clinicians at vessel intersections or bifurcations.

The registration network emits keypoints directly on a shared detection network, and a frequency-guided structure preservation descriptor writes vessel locations, boundaries and skeletons into the descriptor, forming coarse keypoint topology constraints at the descriptor level. The two-step pipeline feeds segmented vessel structures into the detection network, making segmentation quality the performance ceiling; the one-step approach instead lets vessel structure similarity itself carry the constraint. The frequency-guided structure preservation descriptor combines a frequency-aware dynamic kernel generator, a low-frequency aggregator and a high-frequency enhancer; the aggregator locates overall vessel shape, and the enhancer uses a Sobel filter to emphasize ambiguous vessel boundaries and skeletons.

Cross-modal vessel structure similarity learning further refines the keypoint topology constraints so that keypoints from the two modalities land at the same topological positions. Descriptors alone yield only coarse constraints, and keypoint topology remains inconsistent between the two modalities. The method first mines stable keypoints from predicted heatmaps within a single modality, then uses a bidirectional geometric projection mechanism to warp one side's heatmap into the other via the ground-truth transformation and take an element-wise maximum, and finally optimizes descriptors with cross-modal triplet learning.

Evaluation is carried out on the MICRO benchmark: MICRO-pre is split 6:1:3 into training, validation and test sets, while MICRO-pre2intra and MICRO-pre2post are held out entirely from training as independent sets for out-of-distribution generalization. Images are resized to 768×768, and registrations are graded as failed, inaccurate or acceptable, with acceptable requiring a median error below 20 pixels and a maximum error below 50 pixels. The authors state that the MICRO dataset, code and comparable deep registration methods will be released on GitHub.

The intraoperative and postoperative subsets serve only as held-out test sets and take no part in training, with success rates of 0.227 and 0.740 respectively, below the 0.784 of the preoperative subset; the intraoperative subset has only 28.22% average field-of-view overlap and includes surgical instrument interference. The authors plan future theoretical analysis of MRIRG and further investigation of generalization and robustness on other cross-modal image rigid registration tasks.