Subscribe to RSS
Artificial intelligence for polyp characterization: easy as ABCReferring to Rondonotti E et al. p. 14–22
The concept of real-time optical diagnosis, using enhanced imaging techniques to predict histopathology, has been pursued for over a decade. Sound principles have guided this enthusiasm, with the most robust case existing for diminutive (< 5 mm) colorectal polyps, which constitute the most commonly identified and excised lesions for histopathological assessment. Abandoning histopathology, and instead relying on optical diagnosis, represents a paradigm shift, often referred to as a “resect-and-discard” strategy for diminutive adenomas and a “leave-in-situ” strategy for diminutive rectosigmoid hyperplastic polyps. This is highly attractive, considering the significant economic and environment burden of the current approach. Moreover, diminutive lesions have an extremely low risk of harboring invasive neoplasia. Owing to a number of barriers, the uptake of optical diagnosis for diminutive colorectal polyps has been poor. In particular, evidence suggests that optical diagnosis is operator-dependent, with nonexperts unable to consistently achieve the performance thresholds required.
Computer-aided characterization (CADx) of colorectal polyps, leveraging artificial intelligence (AI), has rekindled optimism for the incorporation of optical diagnosis into clinical practice. This has been partly fueled by the exponential growth of AI publications relating to colonoscopy. The challenge lies in separating hype from reality. The majority of published CADx studies are preclinical and subject to potential selection bias, leading to an overestimation of performance. Furthermore, the endoscopist and AI combined final diagnosis is often neglected, despite being most relevant to real-world application. A common statement in the medical AI arena is that the field is high on promise but relatively low on proof. The authors of the study by Rondonotti et al. should be commended for providing much needed high quality prospective evidence .
“AI assistance met the required thresholds for implementation, and crucially did not have a detrimental effect on the endoscopist, including nonexperts.”
In this study, the authors conducted a multicenter prospective study across four endoscopy departments in Italy, including nine expert and nine nonexpert endoscopists. The primary aim was to determine whether high confidence AI-assisted optical diagnosis using a CADx system (CAD-EYE; Fujifilm Co., Tokyo, Japan) could achieve the ≥ 90 % negative predictive value (NPV) for adenomatous diminutive rectosigmoid polyps (DRSPs), according to the Preservation and Incorporation of Valuable endoscopic Innovations (PIVI)-1 threshold . Secondary aims included comparative analyses for endoscopist and AI-alone optical diagnosis performance, whilst finally determining whether the optical diagnosis-based post-polypectomy surveillance achieved the ≥ 90 % agreement rate recommended in the United States Multi-Society Task Force (USMSTF) and European Society of Gastrointestinal Endoscopy (ESGE) guidelines, according to the PIVI-2 threshold.
A sequential diagnostic process was conducted for DRSPs. The first step involved the endoscopist alone making an optical diagnosis according to the Blue-Light Imaging Adenomas Serrated International Classification (low or high confidence); the AI system was then switched on to provide an AI-alone optical diagnosis in the second step; before a final third step was concluded, which involved the endoscopist making a final AI-assisted optical diagnosis (low or high confidence) by combining the first two steps.
AI-assisted optical diagnosis was made with high confidence in 550 DRSPs with performance meeting both the PIVI-1 and PIVI-2 thresholds, achieving an NPV of 91.0 % (95 %CI 87.1 %–93.9 %) for adenomatous pathology and agreeing with post-polypectomy surveillance intervals in 97.4 % (95 %CI 95.7 %–98.9 %) and 92.6 % (95 %CI 90.0 %–95.2 %) according to ESGE and USMSTF guidelines, respectively. There were no significant differences in accuracy parameters when comparing the three sequential diagnostic steps.
These findings are of importance, particularly because this study design is arguably the closest to mirroring real-life practice. AI assistance met the required thresholds for implementation, and crucially did not have a detrimental effect on the endoscopist, including nonexperts. The study also demonstrated a similar accuracy for optical diagnosis when comparing proximal and distal diminutive polyps, in contrast with other published studies that have reported inferior accuracy proximally, where it has been suggested that perhaps visual characteristics may not correlate as well with the histopathology .
Interestingly, the AI-alone performance in this study did not meet the required PIVI-1 threshold. An AI-assisted optical diagnosis was also not feasible in 9 % of polyps owing to unstable predictions or the inability of the system to provide a recordable output. These shortcomings will be readily addressed with improved iterations of the AI software. In addition, the nonexperts in the cohort did not meet the required PIVI-1 threshold, with no significant improvement demonstrated with AI assistance; however, only approximately one-third of polyp optical diagnoses were performed by nonexperts, and the study was not necessarily powered to investigate this. Future studies should focus on nonexperts, where the clearest case for the use of AI assistance exists. It is noteworthy that the NPV for the last 50 DRSPs evaluated by nonexperts did meet the PIVI-1 criteria in this study.
There are few published prospective clinical studies that evaluate CADx in combination with endoscopists. Two recently published studies, using non-magnification narrow-band imaging and white light respectively, suggested that both PIVI thresholds can be met; however, the incremental independent benefit of AI was not strictly evaluated  . Another multicenter international trial evaluating CADx alongside ultramagnification endocytoscopy suggested that CADx assistance did not significantly improve the diagnostic sensitivity for diminutive neoplastic polyps . Standardization of study design, including agreed clinical end points defined by consensus, has recently been highlighted as a key research priority for AI in colonoscopy .
There is now valid optimism that AI assistance will bridge the performance gap between experts and nonexperts; however, clinical effectiveness represents just one barrier to optical diagnosis implementation. The perceived medicolegal risk of a resect-and-discard strategy still needs to be addressed, as the addition of AI assistance could introduce more complexity when one considers accountability for medical error, although conversely it may add some protection for the endoscopist against litigation by acting as a decision support tool alongside photodocumentation . Reimbursement policies need to be established to allow for the emergence of AI use in routine colonoscopy, supported by cost-effectiveness data. The convergence of major trends, such as increasingly resource-constrained healthcare systems and a looming global warming crisis, mean the potential financial and environmental benefits of optical diagnosis implementation need to be revisited urgently. Thankfully, the emergence of AI has recaptured imaginations in the endoscopy community to make optical diagnosis a reality.
Article published online:
26 September 2022
© 2022. Thieme. All rights reserved.
Georg Thieme Verlag KG
Rüdigerstraße 14, 70469 Stuttgart, Germany
- 1 Rondonotti E, Hassan C, Tamanini G. et al. Artificial intelligence-assisted optical diagnosis for the resect-and-discard strategy in clinical practice: the Artificial intelligence BLI Characterization (ABC) study. Endoscopy 2022; 55: 14-22 DOI: 10.1055/a-1852-0330.
- 2 Dayyeh BKA, Abu Dayyeh BK, Thosani N. et al. ASGE Technology Committee systematic review and meta-analysis assessing the ASGE PIVI thresholds for adopting real-time endoscopic assessment of the histology of diminutive colorectal polyps. Gastrointest Endosc 2015; 81: 502.e1-502.e16
- 3 Mori Y, Kudo SE, Misawa M. et al. Real-time use of artificial intelligence in identification of diminutive polyps during colonoscopy: a prospective study. Ann Intern Med 2018; 169: 357-366
- 4 Minegishi Y, Kudo SE, Miyata Y. et al. Comprehensive diagnostic performance of real-time characterization of colorectal lesions using an artificial intelligence-assisted system: a prospective study. Gastroenterology 2022; 163: 323-325.e3
- 5 Hassan C, Balsamo G, Lorenzetti R. et al. Artificial intelligence allows leaving-in-situ colorectal polyps. Clin Gastroenterol Hepatol 2022; DOI: 10.1016/j.cgh.2022.04.045.
- 6 Barua I, Wieszczy P, Kudo SE. et al. Real-time artificial intelligence–based optical diagnosis of neoplastic polyps during colonoscopy. NEJM Evidence 2022; DOI: 10.1056/EVIDoa2200003.
- 7 Ahmad OF, Mori Y, Misawa M. et al. Establishing key research questions for the implementation of artificial intelligence in colonoscopy: a modified Delphi method. Endoscopy 2021; 53: 893-901
- 8 Ahmad OF, Stoyanov D, Lovat LB. Barriers and pitfalls for artificial intelligence in gastroenterology: Ethical and regulatory issues. Tech Innov Gastrointest Endosc 2020; 22: 80-84