CATARACTS: Challenge on automatic tool annotation for cataRACT surgery

Area of Science:

• Medical Image Analysis
• Computer Vision in Surgery

Background:

• Surgical tool detection is crucial for analyzing surgical workflows.
• Existing methods primarily focus on laparoscopic surgery, neglecting common procedures like cataract surgery.
• Cataract surgery is the most frequent surgical procedure globally, necessitating specialized tool detection.

Purpose of the Study:

• To evaluate deep learning-based tool annotation algorithms for cataract surgery.
• To compare the performance of algorithms from the CATARACTS challenge against human expert annotations.
• To identify lessons learned for developing future surgical monitoring tools.

Main Methods:

• Analysis of over nine hours of video data from 50 cataract surgeries.
• Utilized datasets with manual annotations of 21 surgical tools by two experts.
• Evaluation of 14 deep learning solutions submitted to the CATARACTS challenge.

Main Results:

• Deep learning models were the primary approach for tool annotation in the CATARACTS challenge.
• The study provides a thorough evaluation comparing algorithm annotations to human interpretations.
• Differential analysis of submitted solutions offers insights into algorithm strengths and weaknesses.

Conclusions:

• The CATARACTS challenge successfully spurred the development of tool annotation algorithms for cataract surgery.
• Findings will guide the design of more efficient tools for real-time surgical monitoring.
• Improved tool detection can enhance surgical training, report generation, and decision support.