OphMatcher: Uncertainty-aware self-training on ophthalmic surgical videos for anatomy-constrained matching and intraoprative navigation.

Abstract

Surgical navigation plays a crucial role in enhancing precision in ophthalmic surgery, with inter-frame video matching serving as the key enabling technology. However, existing image matchers trained on natural scenes fail to meet the specific demands of surgical environments. Developing a specialized matcher faces two major challenges: the lack of annotated surgical datasets and the dynamic nature of intraoperative scenes. We present OphMatcher, a self-training framework for ophthalmic surgical image matching and intraoperative navigation. The proposed framework generates pseudo-labels by enforcing anatomical segmentation constraints, globally selecting difficult frame pairs, and propagating correspondences focused on navigation-relevant structures. It further models pseudo-label reliability through multi-factor uncertainty estimation and contrastive learning to prevent error accumulation, while directly computing navigation parameters from the predicted correspondences. Comparison results demonstrate the state-of-the-art performance of our method over related methods on both held-in and held-out test sets. Ablation studies show the effectiveness of each component in the self-training pipeline. In navigation evaluation, OphMatcher achieves mean rotation errors of 1.68° for reference frame matching and 1.61° for inter-frame matching, satisfying clinical accuracy requirements. We integrated OphMatcher into a surgical microscope and conducted volunteer studies to demonstrate its real-time operation and robust performance under challenging surgical conditions, highlighting its potential for clinical application.