Work

Data and code associated with the publication: Labels for automated step recognition in cataract surgery for videos in the Cataract-1K dataset

Year: 2026

Type: dataset

Abstract: This data repository contains manual annotations for surgical steps in cataract surgery videos from the Cataract-1K dataset, along with outputs from four deep learning models trained to predict step labels at the frame level, and one model which predicts step segments. Manual Annotations: Expert annotators reviewed surgical videos and marked the temporal boundaries of 13 surgical steps for 386 cases, providing ground-truth labels for model training and evaluation. Model Predictions: Four deep learning architectures (GLSFormer, JEPA, MAEattn, and MAEfull) were applied to all 1,056 videos in the Cataract-1K dataset to generate frame-level step predictions, including predicted labels and class probabilities for each frame sampled at 6-frame intervals (approximately 0.2-second resolution at 30 fps). A fifth model, viTAL, was applied to generate segment-level labels for 875 non-training set videos. The dataset supports research in surgical workflow analysis, automated surgical phase recognition, and the development of AI-assisted tools for ophthalmic surgery. All annotations and predictions are provided in CSV format for broad compatibility with data analysis tools. For access to the original surgical videos, see the Cataract-1K dataset: N. Ghamsarian, Cataract-1K, Synapse (2024); https://doi.org/10.7303/SYN52540135

Source: Research Data Repository, Duke University

DOI: 10.7281/t1zzx1nj