Real-time near infrared artificial intelligence using scalable non-expert crowdsourcing in colorectal surgery

Garrett Skinner,Tina Chen,Gabriel Jentis,Yao Liu,Christopher McCulloh,Alan Harzman,Emily Huang,Matthew Kalady,Peter Kim

DOI: https://doi.org/10.1038/s41746-024-01095-8

IF: 15.2

2024-04-23

npj Digital Medicine

Abstract:Surgical artificial intelligence (AI) has the potential to improve patient safety and clinical outcomes. To date, training such AI models to identify tissue anatomy requires annotations by expensive and rate-limiting surgical domain experts. Herein, we demonstrate and validate a methodology to obtain high quality surgical tissue annotations through crowdsourcing of non-experts, and real-time deployment of multimodal surgical anatomy AI model in colorectal surgery.

health care sciences & services,medical informatics

What problem does this paper attempt to address?

The problem that this paper attempts to solve is how to obtain high - quality surgical tissue annotation data through non - expert crowdsourcing in colorectal surgery to train and deploy artificial intelligence (AI) models for real - time multi - modal surgical anatomy, thereby improving surgical safety and clinical outcomes.

Specifically, current surgical AI models usually require a large amount of video data annotated by domain experts to identify surgical phases, key events, and surgical anatomical structures. However, obtaining these expert - annotated data is both expensive and time - consuming, which limits the development and application of surgical AI models. To this end, researchers have explored a method, that is, using a gamified, continuous - performance - monitored crowdsourcing platform to obtain high - quality surgical tissue annotation data from non - experts for training soft - tissue - segmentation AI models, and verified the effectiveness and accuracy of this method.

Through this method, researchers can not only significantly reduce the time and cost required for expert annotation, but also maintain a high annotation accuracy rate, thereby accelerating the development and clinical application of AI models. In addition, the study also demonstrated the application of this AI model in a real - time multi - modal imaging platform, which can help surgeons more accurately assess tissue perfusion and reduce the occurrence of postoperative complications.