Cataract-LMM Large-Scale Multi-Source Multi-Task Benchmark for Deep Learning in Surgical Video Analysis

Area of Science:

• Ophthalmology
• Computer Science
• Medical Education

Background:

• Developing generalizable deep learning models for computer-assisted surgery requires extensive, varied, and deeply annotated video datasets.
• Current cataract surgery datasets lack the diversity and annotation depth needed for robust AI model training.

Purpose of the Study:

• To introduce a comprehensive dataset of phacoemulsification cataract surgery videos to address limitations in existing resources.
• To facilitate the development of advanced AI models for surgical workflow analysis, scene understanding, and competency-based training.

Main Methods:

• Compiled a dataset of 3,000 phacoemulsification cataract surgery videos from two surgical centers, featuring surgeons of varying expertise.
• Annotated videos with four layers: temporal surgical phases, instrument/structure segmentation, instrument-tissue interaction, and skill scores (ICO-OSCAR, GRASIS).
• Benchmarked deep learning models for workflow recognition, scene segmentation, interaction tracking, and skill assessment; established domain-adaptation baselines.

Main Results:

• The dataset encompasses 3,000 videos with multi-layer annotations, capturing clinical and technical variability.
• Deep learning models were benchmarked on tasks including workflow recognition and automated skill assessment.
• Domain-adaptation baselines were established, demonstrating the dataset's utility for transfer learning.

Conclusions:

• The presented dataset, with its multi-source acquisitions and multi-layer annotations, is crucial for advancing AI in surgical education and analysis.
• Facilitates research into generalizable multi-task models for surgical workflow, scene understanding, and competency assessment.
• Enables the development of AI tools for improved surgical training and performance evaluation.