The SAGES Critical View of Safety Challenge

The application of Computer Vision (CV) and Machine Learning (ML) to minimally invasive surgery promises objective assessment of visual features in surgical video that contribute to surgical decision-making. Future prospects for surgical risk mitigation include enhanced supervision for surgeons and augmented teaching opportunities. Currently, surgical Artificial Intelligence (AI) is limited to research, waiting to be translated into clinical practice. There is a lack of widely validated results, calibration of uncertainty, robustness to domain shifts, and a high computational barrier to enable wide deployment that limits the application of AI to life-saving intraoperative use. 
 
Laparoscopic cholecystectomy, a standardized operation for gallbladder removal, is one of the most frequently performed minimally invasive procedures worldwide. While it has become a benchmark procedure for computational exploration of intraabdominal video data, there is no clinical translation, partly due to the limited generalizability of AI architectures developed almost entirely on localized, homogenous datasets. 
 
The Society of American Gastrointestinal and Endoscopic Surgeons (SAGES) Critical View of Safety (CVS) Challenge is the first international biomedical data challenge from a surgical society, offering a unique infrastructure for global data collection and leveraging multidisciplinary expertise for the standardized assessment of the CVS. The aim of the challenge is to computationally address the detection of CVS, a routinely performed surgical safety measure crucial for minimizing bile duct injuries during cholecystectomy. Despite the high frequency of cases and standardized operative approach for laparoscopic cholecystectomy, there is a risk of significant intra- and post-operative complications, such as common bile duct injuries. Assisting surgeons in achieving and recognizing the CVS will help to improve the safety of laparoscopic cholecystectomy worldwide. 
 
The challenge offers a global and diverse dataset of 1000 laparoscopic cholecystectomy videos, provided by 67 surgeons (data donors) from 53 countries and 6 continents, alongside clinically relevant metadata. This dataset encompasses a wide diversity of patient demographics and procedural quality to reflect the worldwide diversity in patients and surgeons. The data will be released to the global community with the aim of developing models capable of reliably and consistently classifying the CVS with adequate generalizability to the global patient population. By incorporating the perspectives of clinicians, computer scientists, and industry through structured multidisciplinary Advisory Committees (AC), the challenge offers the opportunity for the development of AI suitable for high-stakes surgical settings. 
 
Data acquisition was designed to provide consistent and reliable deidentification of out-of-body images and pseudonymization of metadata. The dataset has been meticulously curated based on standardized protocols composed through expert consensus of the multidisciplinary ACs. The data has been indexed according to demographics -- source location, performing surgeons’ experience level, surgical indication and clinical characteristics in the video (e.g., fluorescence, robotics, intraoperative cholangiogram). Each video was annotated with the CVS and its subcomponents to ensure consistency and reliability in the data. The structured annotation pipeline, rooted in a consensus annotation protocol revised by clinical experts in hepatobiliary surgery, includes proficiency-based training of annotators from multiple countries. The annotation task was the classification of the three CVS Criteria on a video and frame basis.
 
The execution of the CVS Challenge is governed by metrics selected by the multidisciplinary AC to evaluate AI models’ performance in identifying the achievement of the subcomponents and overall CVS. Participants can work with various data splits, enabling them to test the robustness of their algorithms across a heterogeneous dataset rich in clinical and demographic variability.
 
The CVS Challenge sets a new benchmark in collaborative efforts between surgeons and computer scientists, encompassing consensus-based guidelines governing a global data acquisition and curation framework. Additionally, structured annotation curricula based on clinical expert consensus, ensure a homogenous, clinically-meaningful ground truth for model training. The primary aim of the challenge is to ensure uniformity in the computational assessment of the CVS for high clinical value, with the goal to improve safety and outcomes of laparoscopic cholecystectomy. The extensive dataset, rigorous curation protocol, and strategic execution of the challenge, backed by a strong advisory framework, pave the way for transformative developments in surgical practice and patient care as well as opening the possibility for future challenge iterations.