Artificial intelligence (AI) software for image analysis have been developed to assist the endoscopist in his performance of image interpretation. The development has been years in the making and while first applications were generally offline, meaning that the employed AI System used still images were obtained during endoscopy. Analysis could only take place after the endoscopic investigation. 1 With the increasing technical developments in the past decade, it is possible to run an AI system online in parallel to the investigation. Usually AI systems work with an overlay of information superimposed on the endoscopic image either with coloured frames, bounding boxes, or in some cases acoustic signals to raise the attention of the endoscopist towards a relevant finding.
The first studies using approved AI systems unanimously demonstrated an increase of adenoma detection rate (ADR), number of adenomas per patient as well as a reduction in adenoma and neoplasia miss rate that was mainly achieved due to detection of small and diminutive polyps. 2 , 3 , 4 , 5
In some studies, the high number of additional polyps detected with AI could at least in parts be attributed to a low baseline ADR in the control groups. 6
In their paper, Silvia Pecere and colleagues performed an analysis of published data. They included six studies with 67 endoscopists and 2058 patients. 7 The main focus was the performance of participating endoscopists in detection and characterisation of colonic adenomas with AI systems compared to the control groups.
Pooled sensitivity and specificity for adenomatous histology was respectively 84.5% and 83%. PPV and NPV were 89.5% and 75.7%. The AUC was 0.82. Expert endoscopists performed better than non‐experts (sensitivity 90.5% vs. 75.5%) and Eastern endoscopists performed with a higher sensitivity than Western (85% vs. 75.8%).
The results show that, in general, the performance of endoscopists in this setting was not satisfying enough to serve as a gold standard. In their analysis Pecere et al. included studies using still images and studies thar were performed on patients during endoscopic procedures, which limits the translation of conclusions to general practice as these data are difficult to compare and per patient analysis is not possible in all cases. In addition, the methodology and setting of the included studies were slightly different in how and how long the images were presented to the investigators. In general, if in doubt about exact characterisation of a polyp, the endoscopist has the ability to investigate extensively during the colonoscopy or to decide to resect the lesion regardless of histology.
Nevertheless, the actual results are extremely interesting and important as they highlight the importance of the comparator to a new method in a comparative trial. The relative increase of detected adenomas attributed to the AI system may be over‐interpreted in case of weak performance of human investigators. This does not automatically alter the overall performance of the AI system.
In addition, in community‐based studies the quality of optical diagnosis is especially low for diminutive polyps, 8 which may be another reason to strengthen the implementation of AI systems.
In general, AI studies need to be interpreted, knowing that the increase of adenoma detection, meaning finding and exact prediction of histology, has natural limits. The lower limit is the baseline ADR that is mostly influenced by technical factors and the endoscopists performance and the upper limit of adenoma detection is the real number of detectable adenoma in the cohort, thus representing a natural asymptotic border of possible improvement. Indeed, even in studies using only still images the performance of endoscopists was suboptimal and this points to the weakest factor of low adenoma detection and correct prediction of histology the endoscopist.
However, reverting to the results of the current paper where the authors implicate to include only investigators who meet certain quality standards limits into scientific trials while at the same time AI can be used for quality assessment and teaching.
This paper is not meant to lower the scientific or clinical expectations of AI systems in endoscopy but clearly stands out how important it is to carefully interpret, especially early scientific data of new methods. AI remains one of the most important technologies of current time and has a great potential to help increasing the detection of gastrointestinal lesions and to interpret the findings to fill the gap between existing and detected adenoma and may play a role in education of endoscopists.
CONFLICT OF INTEREST
I am author of one of the cited articles. I received speaker fees and a research grant from Fujifilm.
REFERENCES
- 1. Maroulis DE, Iakovidis DK, Karkanis SA, Karras DA. CoLD: a versatile detection system for colorectal lesions in endoscopy video‐frames. Comput Methods Progr Biomed. 2003;70(2):151–66. 10.1016/s0169-2607(02)00007-x [DOI] [PubMed] [Google Scholar]
- 2. Sivananthan A, Nazarian S, Ayaru L, Patel K, Ashrafian H, Darzi A, et al. Does computer‐aided diagnostic endoscopy improve the detection of commonly missed polyps? A meta‐analysis. Clin Endosc. 2022;55(3):355–64. 10.5946/ce.2021.228 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3. Shaukat A, Lichtenstein DR, Somers SC, Chung DC, Perdue DG, Gopal M, et al. Computer‐aided detection improves adenomas per colonoscopy for screening and surveillance colonoscopy: a randomized trial. Gastroenterology. 2022;S0016‐5085(22)00519‐4. 10.1053/j.gastro.2022.05.028 [DOI] [PubMed] [Google Scholar]
- 4. Xu H, Tang RSY, Lam TYT, Zhao G, Lau JYW, Liu Y, et al. Artificial intelligence‐assisted colonoscopy for colorectal cancer screening: a multicenter randomized controlled trial. Clin Gastroenterol Hepatol. 2022:S1542‐3565(22)00673‐5. [DOI] [PubMed] [Google Scholar]
- 5. Wallace MB, Sharma P, Bhandari P, East J, Antonelli G, Lorenzetti R, et al. Impact of artificial intelligence on miss rate of colorectal neoplasia. Gastroenterology. 2022;163(1):295–304.e5. 10.1053/j.gastro.2022.03.007 [DOI] [PubMed] [Google Scholar]
- 6. Wang P, Berzin TM, Glissen Brown JR, Bharadwaj S, Becq A, Xiao X, et al. Real‐time automatic detection system increases colonoscopic polyp and adenoma detection rates: a prospective randomised controlled study. Gut. 2019;68(10):1813–9. 10.1136/gutjnl-2018-317500 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7. Pecere S, Antonelli G, Dinis‐Ribeiro M, Mori Y, Hassan C, Fuccio L, et al. Endoscopists performance in optical diagnosis of colorectal polyps in Artificial Intelligence studies. United Eur Gastroenterol J. 2022. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8. Rees CJ, Rajasekhar PT, Wilson A, Close H, Rutter MD, Saunders BP, et al. Narrow band imaging optical diagnosis of small colorectal polyps in routine clinical practice: the Detect Inspect Characterise Resect and Discard 2 (DISCARD 2) study. Gut. 2017;66(5):887–95. 10.1136/gutjnl-2015-310584 [DOI] [PMC free article] [PubMed] [Google Scholar]