Study: Automatic evaluation of cataract surgery videos for optimized training & competition
Cataracts are the most common cause of blindness worldwide, particularly affecting people in lowand middle-income countries such as India. The low-cost and effective SICS surgical method is preferred in these countries, but is often associated with poorer results due to limited resources and training opportunities. "In addition, the application of AI on this technique has not yet been researched enough," says Dr. Maximilian Wintergerst, head of a working group at the UKB Eye Clinic and principal investigator of the study. While algorithms for AI-supported video analysis of the individual surgical phases have already been developed for the predominant cataract surgery technique in high-income countries, known as phacoemulsification, neither data sets nor algorithms have been available for SICS to date. The new study now makes videos of manual small incision cataract operations publicly available for the first time with the "SICS-105" data set. The data set is based on operations performed on 105 patients at Sankara Eye Hospitals in India.
The study shows that the innovative deep learning model "MS-TCN++", developed in the group of Jürgen Gall at the University of Bonn, can recognize different surgical phases such as preparation of the surgical approaches to the eye and the various surgical steps on the lens with over 85 percent accuracy. "The analysis of surgical phases is important because it enables a quantitative comparison between different surgeons, feedback on identified critical steps and the detection of deviations from surgical protocols. It is therefore the first step towards automatic assessment of surgical quality," says Dr. Kaushik Murali, president of medical administration at the Sankara Eye Foundation India. The transdisciplinary approach that enables such progress is also represented by Simon Mueller, first author of the study: After finishing a MSc degree in Life Science Informatics at the University of Bonn, he is now studying medicine in Maastricht; in parallel, he works on a PhD project in close collaboration between the Computer Science Department in Bonn and the UKB Eye Clinic.
SICS-155 Challenge: New milestone for education and research
The research consortium is currently calling for an AI competition to analyze such surgical videos. To this end, the research team has expanded the study’s first public data set for SICS with surgical videos and hand-marked (annotated) surgical phases to a total of 155 operations with 18 different phases. The software for annotation has been developed by researchers at Microsoft Research India and Sankara Eye Foundation. Subsequently, the annotations have been created by ophthalmologists at the Sankara Eye Foundation. "With the ’SICS-155 Challenge’, we are inviting international teams to test their AI algorithms for phase recognition in 155 SICS operations," says Thomas Schultz from the b-it and Institute of Computer Science at the University of Bonn and the Lamarr Institute for Machine Learning and Artificial Intelligence. He is also a member of the Transdisciplinary Research Areas (TRA) "Modeling" and "Life and Health" at the University of Bonn. The international research team expects participants to submit an algorithm for predicting surgical phases based on the video data provided and to write a short paper on their approach. "With the competition, we want to accelerate progress in the automatic analysis of surgical videos from middleand low-income countries and thus improve the training of surgeons and cataract surgery outcomes in the long term," says Wintergerst.
Development of algorithms for instrument and complication detection follows
In addition to the "SICS-155 Challenge", as the next step, computer scientists at Microsoft Research India and University of Bonn are developing algorithms for automated surgical instrument and complication detection, which will further advance automated surgical video analysis for SICS.
