Jin Sol Oh, MD1, Jennifer A Minneman, MD1, Anne P Ehlers, MD, MPH1, Shanley B Deal, MD2, Adnan A Alseidi, MD2, Andrew S Wright1. 1University of Washington, 2Virginia Mason Medical Center
Introduction: Laparoscopic cholecystectomy (LC) is performed over 750,000 times each year in the U.S. with 0.3% risk of bile duct injury. The SAGES Safe Cholecystectomy Program advocates implementing the Critical View of Safety (CVS) method to decrease the risk of bile duct injury. Many surgeons still do not routinely obtain the CVS and may not recognize when it has not been achieved. We hypothesize that machine learning algorithms can be used to construct a decision support tool to assist in recognition of the CVS.
Methods: For algorithm development, 220 de-identified videos of LC were collected. Still images were captured from the video, immediately prior to applying clips and dividing the cystic structures. The images were manually rated on a 6-point CVS scale, using previously published scoring criteria. We developed two algorithm models with theGoogle AutoML Vision platform, with which users can build machine learning algorithm models using labeled images. For the first model, the images were labeled as “good” (CVS score >4), “medium” (CVS score 2-4), and “poor” dissection (CVS score <2). For the second model, they were labeled as “adequate” (CVS score ≥5) or “inadequate” dissection (CVS score <5). The algorithm models were evaluated at a score threshold of 0.5, at which the model predicts the categories with 50% confidence.
Results: There were a total of 292 images. The first model was trained with 60 “good”, 86 “medium”, and 151 “poor” images. The algorithm had an area under the receiver operating curve (AUC) of 0.672. The positive predictive value (PPV) and sensitivity were 62.9% and 59.5%. For the second model, only 17 images were “adequate”. The algorithm model had an AUC of 0.831, and the PPV and sensitivity were 62.5% and 62.5%. Once trained, the algorithm model can be tested on any LC images. (Fig. 1)
Conclusions: We developed two machine learning algorithm models to assess the CVS. Accuracy will likely improve with further model refinement using larger image data sets. This algorithm model may serve as future clinical decision support tool in the operating room.
Presented at the SAGES 2017 Annual Meeting in Houston, TX.
Abstract ID: 95608
Program Number: S068
Presentation Session: Residents and Fellows Session
Presentation Type: ResFel