Post-EMR Scar Detection During Endoscopy

Abstract

Introduction: Endomucosal resection (EMR) is a critical tool for large colon polyp resection. Because EMR for larger polyps is often performed in a piecemeal fashion, follow-up surveillance colonoscopy is typically recommended to evaluate and treat any small areas of polyp recurrence (PMID 37414440). The first task during post-EMR surveillance colonoscopy is to detect the area of scar, particularly if a prior tattoo is not placed or has faded. The second task is to distinguish scar from recurrent adenoma (ref1 and also PMID 31494134). Detecting scars and evaluating for polyp recurrence are both challenging visual tasks which may be aided by artificial intelligence tools. Our overarching goal is to develop a combined CADe/CADx tool, which both detects the location of post-EMR scars, and provides a prediction regarding the presence or absence of recurrent adenoma, in order to improve the efficiency of surveillance procedures, and to potentially reduce the need for biopsy as the scar site. The aim of the present study is to address the first task, by developing a computer vision tool to detect EMR scars during surveillance colonoscopy. Methods: The dataset used to train and evaluate model performance is based on procedures at Beth Israel Deaconess Medical Center recorded from January 2022 until October 2023. Recordings were performed using Olympus endoscopes and Olympus and Virgo recording systems. To create the models, the data were annotated using the open-source annotation tool CVAT and converted into YOLOv5 format. Scars were annotated using bounding boxes. Annotations were performed by medical students and checked by experienced endoscopists. The dataset was randomly split 80/10/10 for training, validation, and testing. The architecture used for the model, YOLOv5l, was trained for 20 epochs on the training data then evaluated on testing data. The model reported mAP (mean average precision), precision, recall, and F1 score. Results: The scar detection model was tested on a held-out testing dataset and was able to achieve a mAP of 0.973, an F1 score of 0.94, a precision of 0.944, and a recall of 0.94 using a confidence threshold of 0.5. An example of a correct detection of an EMR scar is shown in Figure 1. Conclusions: Using annotated videos of EMR scars, we were able to build a model that accurately detects post-resection scars during surveillance colonoscopy. In the next phase, using more detailed labeling of the pathology of the scar (normal or adenomatous), we intend to further develop the model to accurately distinguish between flat/nodular scar tissue vs. recurrent adenoma under white light and narrow band imaging.