Thriving with Artificial Intelligence, AI Tips and tricks in optimising AI-guided colonoscopy in real-time polyp identification and characterisation These are our disclosures Colonoscopy has long been considered the gold standard technique employed for the screening and diagnosing of colorectal malignancies With the widespread utilisation of endoscopy worldwide Artificial Intelligence-guided endoscopy systems have been introduced to boost efficacy and accuracy of lesion detection. Studies have shown that such integration of AI technology into clinical endoscopic work has indeed proved beneficial. With both improved polyp and adenoma detection rates, decreased burden of polyp detection on individual endoscopists alone, as well as early adenoma detection resulting in perverted decreased rates of colorectal cancer diagnosis. However despite the advent of AI use in colonoscopy there has not yet been the release of any standardised instructional videos on how best to optimise and capitalise on such AI programs. This unmet clinical need for tips on troubleshooting limitations of the AI system thus allows for further discussion on ways to help endoscopists master nuances in maximising the AI algorithm in clinical practice The AI algorithm in question was notably trained and programmed to detect and subsequently characterise polyps and adenomas. Sustained detections, such as that depicted in the photo, would be highlighted within a green box with one out of the three possible predicted characterisations labelled after a period of analysis. A total of 428 colonoscopies from the endoscopy suite in a single Singaporean institution over a period of 6 months were analysed to better understand the common pitfalls and difficulties encountered during such AI-guided colonoscopy sessions in order to evaluate ways to further improve and navigate the system's limitations As previously described, the endoscopist focuses on the desired lesion. Appearance of the green box suggests sustained detection of the polyp by the system, and within a few seconds of stable field of vision, the system generates a real-time characterisation of the polyp, with one of the three predictions appearing in the label adjacent to the lesion. In this video, a colonic adenoma was accurately identified. A similar process is seen here, with the detection of a non-adenomatous lesion instead The third potential characterisation by the system would be that of a null prediction, in which the AI system is able to detect a polyp, but unable to confidently classify it as either of the former categories, as seen in this video Such null predictions were a common occurrence amongst the polypectomies in the scope videos assessed Upon further evaluation, these null predictions were often tied to a few common modifiable factors, namely inadequate analysis duration, non-stable field of vision during analysis, as well as poor bowel preparation Malignant lesions also often yielded null prediction, which was somewhat unsurprising in view of the algorithm's focus on detection and characterisation of mainly adenomas This led us to the clinical question of how then to navigate these challenges in order to best maximise and utilise the AI technology As per any endoscopic procedure, optimising of mucosal visualisation is a basic, effective way in maximising polyp detection with or without AI guidance. Good bowel clearance reduces the chance of residual fecal matter obscuring small lesions especially, hindering both detection and analysis This can be overcome to an extent with the help of manual water wash and sometimes the administration of antimotility agents. Insufficient time duration for analysis was also often accorded after the detection of a polyp Although challenging at times, maintaining of a sustained steady clear visualisation of the desired lesion allows for better analysis. Though this may marginally slow down the experienced endoscopist, this time can instead be used by the procedurist to make his or her own assessment of the lesion Tackling null predictions proved tricky However, repeat analysis of the lesion was noted to most consistently be a good mitigating measure With removal of the polyp from the field of vision, followed by a subsequent refocusing on the same lesion as seen in this video, reanalysis by the AI algorithm is allowed, with repeat opportunity for generating of a predicted characterisation. Another challenge noted in the use of the above AI system was its limited ability to detect or predict malignant lesions This was fairly unsurprising, given the algorithm was neither trained for the identification of malignant lesions, neither was it programmed to distinguish between grades of dysplasia. As such, the responsibility falls on the individual endoscopist to recognise suspicious lesions. Reliance on AI programs for this would be inappropriate and unwise As AI-guided colonoscopy grows to become more accurate and reliable, there is potential for a shift in the focus of polypectomies to advocating for more targeted polypectomies. This promises benefits of reduced overall procedural risks and overall minimisation of healthcare costs. Without a standardised guide in ways to troubleshoot common pitfalls and limitations of this particular AI system, there is much room for endoscopists worldwide to share and exchange helpful techniques in order to fully optimise and improve existing AI algorithms In conclusion, AI technology is a useful adjunct in colonoscopy. The aforementioned measures can be taken by endoscopists to troubleshoot and navigate the modifiable limitations of the algorithm With reliable polyprediction technology, there is potential for reduction in non-essential polypectomies of non-pre-cancerous lesions. However, endoscopists must not become reliant on AI. Clinical recognition remains key in the identification of suspicious lesions Thank you

Artificial Intelligence

Colonoscopy

Lesion detection rates

Endoscopy systems

Colorectal cancer diagnosis