Award: Presidential Poster Award
Susan Y. Quan, MD1, Shai Friedland, MD2, Hamed Pirsiavash, PhD3, Ravindra Kompella, MS4, Vineet Sachdev, MS5
1VA Palo Alto Health Care System / Stanford, Palo Alto, CA; 2Stanford University and VA Palo Alto, Stanford, CA; 3University of Maryland, Baltimore, MD; 4Endovigilant, LLC, Hyderabad, Telangana, India; 5EndoVigilant, LLC, Millersville, MD
Introduction: Artificial intelligence (AI) is increasingly being used as an aid to physicians in multiple areas of medicine. Computer aided detection (CAD) has the potential to assist endoscopists in detection of colon polyps and flat lesions during colonoscopy. Real-time detection of polyps can potentially increase adenoma detection rates and decrease adenoma miss rates, as computerized algorithms are able to analyze the entire field of view more rapidly than humanly possible and do not suffer from fatigue or distraction.
Methods: We designed a deep learning based CAD system that is able to process 30 endoscopy video frames per second in real time and highlight one or more polyps that are detected anywhere in the field of view. Our deep learning algorithm was trained on 26000 images extracted from hundreds of procedure recordings. The detection performance of the CAD system was measured on 100 colonoscopies. To compare how early the CAD system detects relative to endoscopists, 25 different colonoscopy video clips of varying length (30-180 sec) that included a single polyp were reviewed by 5 experienced endoscopists. The endoscopist pressed a button on the computer keyboard as soon as the polyp was visualized, and the time was recorded. This was compared to the time at which the CAD system detected the polyp.
Results: In the 100 clinical colonoscopy cases, a total of 250 polyps were found and removed in 86 patients (14 patients had no polyps detected). There were 171 tubular adenomas, 54 hyperplastic polyps/mucosal excrescences, 11 inflammatory polyps, 6 villous adenomas, 6 sessile serrated polyps and 2 traditional serrated adenomas. The CAD system detected 247/250 polyps (98.8%). The 3 polyps that were not detected were: one 6mm serrated polyp, one 4mm tubular adenoma and one 3mm tubular adenoma. Separately, as shown in Figure 1, the early detection experimental analysis revealed that more than 49% of the polyps were detected earlier by the CAD system by 2 seconds or more. CAD system typically detected much earlier when physicians initially missed brief transient visibility (< 500ms) of the polyps.
Discussion: Our deep learning based CAD system is capable of detecting 99% of polyps that are found during routine colonoscopy. The system detects polyps significantly faster than physicians. Future studies will address whether this translates into higher adenoma detection rates or reduced miss rates in clinical practice.
Citation: Susan Y. Quan, MD; Shai Friedland, MD; Hamed Pirsiavash, PhD; Ravindra Kompella, MS; Vineet Sachdev, MS. P0244 - ARTIFICIAL INTELLIGENCE BASED COMPUTER AIDED DETECTION SYSTEM RELIABLY DETECTS POLYPS EARLIER THAN PHYSICIANS DURING COLONOSCOPY. Program No. P0244. ACG 2019 Annual Scientific Meeting Abstracts. San Antonio, Texas: American College of Gastroenterology.