Discerning surgical phases in laparoscopic surgeries is an essential step for computer-assisted surgery (CAS) applications. Several past studies utilize computer vision in the field of surgical workflow analysis. Despite the promising results, the question of generalization of such studies remains unclear and makes it hard for clinicians to trust and adopt related applications in their practice.
In this work, we present the results of a study that includes 587 lap-cholecystectomy surgical videos which were manually annotated with workflow surgical phases. This dataset, curated from five different medical centers and performed by 49 different surgeons, is an order of magnitude larger than all prior studies, it is captured from a much larger variety of institutions and surgeons and was manually annotated with surgical workflow phases.
This unique dataset makes it possible to train a state-of-the-art deep spatio-temporal model which focuses on short and long term surgical context. We then assess both the likely asymptotic performance of such a workflow phase-detection system and evaluate how dataset size affects the generalization of the system to new examples from unseen surgeons and/or institutions.
Our results suggest that a model trained on several hundred samples can generalize well and achieve above 90% frame-level accuracy while maintaining unbiased results toward medical centers or surgeons. Moreover, given a set of unseen videos from a new medical center, we show that by fine-tuning on a relatively small number of new samples, our model can converge to the same performance.
About Dotan Asselmann
Dotan jumpstarted his professional career while serving as the computer vision team leader of Israel's 81 military intelligence unit. After six years of service, he co-founded EyeOnn, a startup tackling drowning events in pools by leveraging computer vision. On a mission to make the world a bit better, Dotan co-founded theator, and serves as the company's CTO. He also heads Israel’s largest computer vision meet-up group (nearly 3000 members and growing).