Subscribe to RSS
Computer-aided detection is here: will computer-aided quality algorithms soon follow?Referring to Yao L et al. p. 757–768
In the auditorium following a presentation on how computers can automatically highlight polyps with computer-aided detection algorithms, an endoscopist stood up to ask a question: “This will find polyps that are in the field of view,” he noted. “But what about the polyps that weren’t exposed?”
“The results from Yao et al. add to evidence that withdrawal technique and mucosal exposure are critical to polyp detection.”
He was right. Polyps can be missed if they are hidden in mucosa the endoscopist does not expose – the proximal side of the ileocecal valve, the hepatic flexure, the deep rectal valves. These colonoscopy blind spots assuredly contribute to the one in five adenomas missed on tandem studies, and to the missed adenomas and sessile serrated polyps that are the major source of postcolonoscopy colorectal cancers. Computer-aided detection that places a box around a polyp in view or uses another visual alert has no chance of detecting those polyps that are in unexamined mucosa. To fulfill its promise as the “tireless expert observer in the room,” artificial intelligence (AI) must also address these missed lesions.
Computer-aided detection is here now . It follows feverish work on multiple continents by scientists and clinicians. A plug-and-play algorithm on the market in Europe and the United States (Medtronic) uses artificial intelligence to place a green bounding box on the video feed around detected polyps. The endoscope manufacturers (Olympus, Pentax, Fuji) all have proprietary algorithms approved for use in Europe. Olympus and Pentax use the familiar bounding box, whereas Fuji highlights the shape of the polyp on a separate panel to the right of the video feed – yellow for neoplastic, green for hyperplastic.
Will computer-aided quality assurance algorithms that reduce blind spots soon follow? AI that addresses the concern of the questioning endoscopist at the presentation is also in development and comes in different forms. Our group presented preclinical studies on AI that reconstructs the colon video into a three-dimensional rendering of the colon and highlights blind spots in real time . Even when examining video from an endoscopist with a high adenoma detection rate (ADR), the AI found that 20 % of the mucosa was in fact within blind spots. That 20 % proportion corresponds eerily to the one in five adenomas missed in prior studies.
In this issue of Endoscopy, Yao et al. from Wuhan, China, report the results of a randomized trial that used a different approach: the colonoscopy “speedometer” . To the right of the colonoscopy video, the speedometer measures withdrawal speed using a convolutional neural network and alerts endoscopists to rapid withdrawal rates, and ultimately for the study at hand, calculated the proportion of frames that were “over-speed.” We assume there were no fines for speeding, but the real-time feedback allowed endoscopists to slow down if they chose . The theoretical framework behind the speedometer is based on studies that associated slower withdrawal times to increases in adenoma detection.
The results were intriguing. Aiding withdrawal technique, they suggest, may be even more important than aiding polyp recognition. When 1076 colonoscopy patients were randomized to a speedometer group, a computer-aided detection group, a combination group that received both speedometer and computer-aided detection information, or a control group, all AI groups achieved higher ADRs than the control group, and the combination group had the highest ADR. However, the group with the speedometer alone had a higher ADR than the computer-aided detection alone group, and adding computer-aided detection to the speedometer did not lead to a statistically significant higher ADR than the speedometer alone.
The results from Yao et al. add to evidence that withdrawal technique and mucosal exposure are critical to polyp detection. Of course, robust studies have linked longer withdrawal times, used as a proxy for mucosal exposure, to higher adenoma detection and fewer postcolonoscopy cancers . However, more direct and granular data exist. In a large trial in 2004, 1200 adults underwent same-day computed tomography colonography and colonoscopy, and endoscopists found 21 adenomas measuring 6–17 mm only after the colonography results were unblinded . Half of the missed medium-sized and large neoplasms were on the proximal side of a haustral fold, an area we know commonly holds blind spots. Four of five rectal polyps were within 10 cm of the anal verge, at a time when rectal retroflexion was not standard. Both findings point to colonoscopy blind spots as major factors in the missed polyps.
In an even earlier, single-author study, Doug Rex selected videos of colonoscopy withdrawals by two endoscopists – one with an adenoma miss rate that was double the other – and presented them to expert colonoscopists . All experts judged that withdrawal quality in all the studied aspects (evaluating the proximal sides of folds and valves, cleaning, time spent viewing) was superior in videos from the endoscopist who was a high adenoma detector. On a different note, among the mechanical aids to colonoscopy, a distal attachment cuff that opens up the colon folds for viewing has arguably been the most successful .
When good data exist, physicians owe it to their patients to embrace technology in a wise way that will lead to better outcomes. However, as new technologies throughout history have elicited anxieties, such as the train, the automatic wheat thresher, and the smartphone, anxieties over adoption of AI in colonoscopy will arise, and I share with other endoscopists some of these concerns. Will polyp recognition software help us become better at this important skill, or just lazier, as the computer will now “do it” for us? Will the alerts for blind spots, polyps, and speeding become information overload?
AI for colonoscopy is here. The currently available algorithms assume that endoscopists need help with recognizing and diagnosing polyps to guide management (leave in or resect fully) in real time. This is true, but it is not enough. We need real-time feedback on mucosal exposure as well. Addressing both, technology can realize its potential to aid colonoscopists in their quest to make colorectal cancer after colonoscopy a never event.
Article published online:
25 March 2022
© 2022. Thieme. All rights reserved.
Georg Thieme Verlag KG
Rüdigerstraße 14, 70469 Stuttgart, Germany
- 1 Koo CS, Dolgunov D, Koh CJ. Key tips for using computer-aided diagnosis in colonoscopy – observations from two different platforms. Endoscopy 2021; 53 DOI: 10.1055/a-1701-6201.
- 2 McGill SK, Rosenman J, Wang R. et al. Artificial intelligence identifies and quantifies colonoscopy blind spots. Endoscopy 2021; 53: 1284-1286
- 3 Yao L, Zhang L, Liu J. et al. Effect of an artificial intelligence-based quality improvement system on efficacy of a computer-aided detection system in colonoscopy: a four-group parallel study. Endoscopy 2022; 54: 757-768
- 4 Gong D, Wu L, Zhang J. et al. Detection of colorectal adenomas with a real-time computer-aided system (ENDOANGEL): a randomised controlled study. Lancet Gastroenterol Hepatol 2020; 5: 352-361
- 5 Shaukat A, Rector TS, Church TR. et al. Longer withdrawal time is associated with a reduced incidence of interval cancer after screening colonoscopy. Gastroenterology 2015; 149: 952-957
- 6 Pickhardt PJ, Nugent PA, Mysliwiec PA. et al. Location of adenomas missed by optical colonoscopy. Ann Intern Med 2004; 141: 352-359
- 7 Rex DK. Colonoscopic withdrawal technique is associated with adenoma miss rates. Gastrointest Endosc 2000; 51: 33-36
- 8 Rex DK, Repici A, Gross SA. et al. High-definition colonoscopy versus Endocuff versus EndoRings versus full-spectrum endoscopy for adenoma detection at colonoscopy: a multicenter randomized trial. Gastrointest Endosc 2018; 88: 335-344