Real-time computer-aided polyp detection and localization for clinical applications

Abstract

Gastroenterology is one among many recent medical fields which can benefit significantly from the incorporation of machine learning and computer vision to improve quality of care. We set out to architect and build Skout[superscript TM], a product to aid physicians by helping to localize adenomous polyps during colonoscopies in order to reduce risk of colorectal cancer. We use state-of-the-art computer vision advances, specifically a combination of classification and object detection models as well as object tracking, to localize polyps in real-time in HD-quality colonoscopy streams. While many other companies and research entities are already working on similar products, we aim to improve upon their work with a focus on building a large dataset for training and validation and building a product and system of evaluation around a quality user experience and clinical relevance. We demonstrate success in the current stages of development thus far in achieving clinical relevance through non-clinical metrics indicating low time-to-detection and high sustained detection of polyps at a relatively low FPR and through our pilot study, which demonstrates the promise of our product in boosting performance metrics in effectively all relevant areas without significant negative impact, such as increasing ADR from 40.6% in the 283-patient control group to 54.2% in our 83-patient study group (p = 0.028), for which physicians performed the procedure with the help of Skout[superscript TM]. This thesis was completed at Iterative Scopes, a Boston startup working on bringing precision medicine and technology to gastroenterology.