Eyeris: Eye Surgery Refinement Through Intelligent Simulation

Objective and reproducible assessment of cataract surgical performance remains a major challenge in ophthalmic education. Current training models rely heavily on expert observation and subjective scoring, which limits standardization, scalability, and timely feedback. We recently developed eyeRIS (Eye Surgery Refinement through Intelligent Simulation), an AI-driven platform designed to provide automated, phase-aware assessment of cataract surgery using simulation-based video data. In this study, we investigated whether vision-transformer–based video embeddings capture clinically meaningful information by examining their association with defined phases of cataract surgery. 

Simulated cataract surgery videos were recorded in a controlled laboratory environment (MUHC Ophthalmology Simulation Laboratory) using artificial eye models and standard phacoemulsification equipment and instrumentation. Temporal surgical phase annotations were performed in collaboration with experienced cataract surgeons to ensure clinical relevance.

Fifteen fully annotated simulated cataract surgery videos were segmented into short temporal clips. Frame-level visual features were extracted using an EfficientNet-B0 backbone. Dimensionality reduction with principal component analysis (PCA) preserved 95% of variance prior to unsupervised clustering, allowing exploration of latent visual structure without predefined phase labels. In parallel, supervised phase classification was performed using clip-level embeddings and logistic regression. Model performance was evaluated using held-out validation and leave-one-video-out cross-validation. Phase-specific performance was interpreted in relation to ICO-OSCAR surgical domains.

Across 15 fully annotated simulation videos, supervised models demonstrated reliable automated recognition of cataract surgery workflow and phase-level characterization. In the unsupervised analysis, clustering of PCA-reduced embeddings (K = 19, selected by silhouette score) revealed a strong association between specific clusters and the phacoemulsification phase, suggesting that this technically demanding portion of surgery is robustly encoded by the learned visual representations. Beyond phase correspondence, observed cluster transitions and inter-trial variability suggest that embedding-based features capture differences in surgical execution that may be relevant for objective skill differentiation.

Vision-transformer–based video embeddings capture clinically meaningful information related to cataract surgery phases in a simulated setting. Combining supervised phase recognition with self-supervised clustering reveals latent structure associated with both surgical workflow and technique, without requiring exhaustive expert annotation. This approach addresses a key limitation of many AI-based assessment systems and provides a scalable foundation for objective, competency-based feedback in cataract surgery training. Future work will focus on dataset expansion, external validation, and translation to real-world surgical training environments.