Artificial Intelligence in Colonoscopy – A Brave New World
Written by E. Gibbons, B Hall, Department of Gastroenterology, Connolly Hospital, Blanchardstown
Abstract: Artificial intelligence (AI) has shown potential to improve polyp detection in colonoscopy. Advanced machine learning models are highly effective at image recognition, hence their applicability to polyp detection. In 2018, the World Endoscopy Organisation published a consensus statement on postcolonoscopy cancer (PCCRC), identifying PCCRC rates as an important performance indicator for an endoscopy service.
In one large community-based US study with 1000 identified PCCRCs, almost half had potentially modifiable factors related to polyp surveillance or removal, and examination completeness. These modifiable factors represent areas where AI technology can assist in improving outcomes. The three leading endoscopy manufacturers are currently developing AI systems which will be integrated into future endoscopy processing stacks.
The authors have recently trialled the Medtronicdeveloped GI-Genius system. The combination of advances in AI technologies, commercial interest and a growing body of supporting evidence suggests that the use of AI to augment day to day endoscopy will become a reality in the near future.
Two physicians in the United States, Dr William Wolff and Dr Hiromi Shinya performed the first colonic polypectomy in 1969. Despite advances in operator technique and optic-enhancing technologies since that time, up to 25% of adenomatous polyps can still be missed at colonoscopy. 1 The adenoma-carcinoma sequence describes the transformation of normal colorectal epithelium to adenomatous polyp to invasive cancer. This temporal sequence of events offers an opportunity for endoscopic intervention at a precancerous stage. Despite this, colorectal cancer (CRC) is the third most common cancer in Ireland, accounting for 11% of all cancer related deaths. 2 Artificial intelligence (AI), in particular deep machine learning, is showing potential to improve polyp detection in colonoscopy.
AI is a broad term which refers to the use of computer algorithms to augment human intelligence. Deep learning, a subset of the machine learning field within AI, is a multi-layered mathematical algorithm which creates networks known as artificial neural networks. Unlike traditional machine learning, deep learning can progressively learn independently from human intervention, therefore enabling improved accuracy over time without manually rebuilding the algorithm. Advanced machine learning models such as convolutional neural networks are highly effective at image recognition, hence their applicability to polyp detection.
In 2018, the World Endoscopy Organisation published a consensus statement on postcolonoscopy cancer (PCCRC), identifying PCCRC rates as an important performance indicator for an endoscopy service. 3 In one large community-based US study with 1000 identified PCCRCs, almost half had potentially modifiable factors related to polyp surveillance or removal, and examination completeness. 4 Individual endoscopist adenoma detection rates (ADR) are an independent predictor of the risk of interval cancer. 5 ADR is defined as the proportion of colonoscopies where one or more adenomas are detected. With every 1% increase in an endoscopist’s ADR, their risk of an interval cancer decreases by 3%. 6 These modifiable factors represent areas where AI technology can assist in improving outcomes.
The three leading endoscopy manufacturers are currently developing computer-aided diagnostic (CADx) and computeraided detection (CADe) systems which will be integrated into future endoscopy processing stacks. The authors have recently trialled the Medtronic-developed GI-Genius system. An 11% increase in the polyp detection rate was captured during this period.
There are a growing number of studies demonstrating the potential of these AI systems in colonoscopy. Most systems have been evaluated by using either images or video of real procedures. Multiple systems have the ability to be used in real time, yet only a small number of studies have been performed during real time colonoscopy to date. One of these studies showed an improvement in ADRs with 29.1% in those with a CADe system versus 20.3% in those without. The increase seems to have been due to a higher number of diminutive adenomas and hyperplastic polyps, however, and a cost benefit analysis of the increased yield of diminutive polyps is yet to be fully explored. 7
An Italian team using GI-Genius by Medtronic, randomly assigned 685 individuals to undergo colonoscopy either with CADe or without and found adenomas five mm or smaller were detected in a significantly higher proportion of examinations in the CADe group (33.7%) than in the control group (26.5%). This study found no increase in withdrawal time despite the extra polyps identified, an important consideration for service provision. 8
There is also growing evidence to support the use of CADx systems in a ‘resect and discard’ strategy for tiny, hyperplastic rectal polyps, with studies showing a negative predictive value ranging from 95% to 98% for diminutive rectosigmoid adenomas. 9 Despite the increasing body of evidence, large randomised controlled trials are needed. A systematic review and meta-analysis of the use of AI systems in detection of polyps in Endoscopy this year identified only five randomised trials for analysis. 10
The combination of advances in AI technologies, commercial interest and a growing body of supporting evidence suggests that the use of AI to augment day to day endoscopy will become a reality in the near future. As this happens it will raise a number of key ethical and regulatory issues, including data protection and ownership of missed diagnoses. The relative novelty of this field also poses a challenge in terms of interpreting data presented.
Widespread understanding of the technical aspects of AI doesn’t currently exist among medical professionals, while developers of this technology can be unfamiliar with clinical relevance and implications for daily practise. Important areas that need further investigation include the effect CAD systems will have on length of procedures, their impact on surveillance, and cost of procedure itself, as well as the cost of additional histopathology. For GI endoscopy to really harness the power of this technology, a multidisciplinary approach will be necessary to ensure progress continues and focus remains on clinically relevant patient outcomes.
Declarations:
Funding: Not applicable
Conflicts of interest/ Competing interests: Not applicable
Availability of data and material: Not applicable
Code availability: Not applicable
If you would like to receive an HPN magazine every month, complete the form below to be added to the mailing list. Only those who meet the expected criteria for readership will be added to the list. SUBSCRIBE
Follow us: