Work

Labels for automated step recognition in cataract surgery for videos in the Cataract-1K dataset

Year: 2026

Type: data-paper

Abstract: We provide surgical step annotations for all 1,000 videos in the publicly available Cataract-1K dataset of phacoemulsification cataract surgery. We manually labeled 386 videos using an annotation protocol with explicit boundary criteria for 10 surgical steps. Manual annotations demonstrated excellent inter-annotator reliability (intraclass correlation coefficient 0.994–1.000; mean intersection over union 0.980–0.993). Using the manually annotated videos, we trained five state-of-the-art deep learning models for surgical step recognition and localization and generated pseudo-labels for all 1,000 videos. We release the manual annotations, model-predicted pseudo-labels, and associated prediction probabilities through a publicly accessible repository. These data products, alongside our annotation protocol and statistical evaluation methodology, serve as a resource for developing context-aware systems for cataract surgery and as a guidance for annotation of steps and statistical evaluation of large surgical video datasets.

Source: Scientific Data

DOI: 10.1038/s41597-026-08528-x