Endoscopy 2021; 53(12): 1227-1228
DOI: 10.1055/a-1408-3489

Artificial intelligence for polyp characterization: a challenging road ahead

Referring to van der Zander QEW et al. p. 1219–1226
Emanuele Rondonotti
Gastroenterology Unit, Valduce Hospital, Como, Italy
› Author Affiliations

Replacing histology with optical biopsy is one of the most challenging tasks in gastrointestinal endoscopy. We all agree that histology represents one of the most expensive items in colonoscopy. In that case, we need to question how productive or cost effective it is to send an innocent-looking lesion, such as a hyperplastic polyp, for further evaluation, wasting valuable resources without any clinical benefit.

To facilitate the implementation of the optical-based strategy, a favorable scenario was set, as follows. 1) A decision was taken to focus primarily on diminutive polyps; these lesions represent up to 80 % of all polyps and have a negligible risk of harboring advanced features, minimizing the uncertainty that may undermine the shift from histology to optical biopsy [1]. 2) Experts provided simplified lesion classifications taking into account superficial features (e. g. vascular pattern, pits distribution) under blue-light chromoendoscopy [2] [3]. 3) Structured training/retraining programs for endoscopists were validated. 4) Scientific societies recommended thresholds for maximizing the benefit/risk ratio [1] [4]. Despite these advances, optical biopsy is still far from replacing histology as an acceptable clinical standard.

There are several reasons for this lack of progress. First and foremost, generalizability of classification-based optical biopsy results in the community setting is poor; the DISCARD 2 trial [5] showed the overall accuracy in polyp characterization to be well below 90 %. Second, dogmatic perceptions regarding the transformation of every removed tissue piece to a histology specimen remains prevalent. Third, there are issues associated with reimbursement, as well as medicolegal concerns. Fourth, optical biopsy requires consistent, precise, and detailed photodocumentation. Last but not least, there is understandable concern among many stakeholders of replacing a well-accepted objective and clinical standard with another potentially inferior practice, just for the sake of macroeconomic savings.

“Thus, instead of abolishing the need for training and education, a high dedicated competence in polyp characterization without AI assistance becomes a prerequisite for anyone wishing to use AI in clinical practice.”

Therefore, any new system with the potential to provide a reliable polyp characterization, by increasing the endoscopist’s performance and decreasing interobserver variation, undeniably stirs great enthusiasm and creates even greater expectations. As polyp characterization involves quick processing of several images, search for specific visual clues, and recognition of well-defined visual patterns, it represents a fertile ground for dedicated convolutional neural network (CNN)-based artificial intelligence (AI) systems. In this issue of Endoscopy, van der Zander et al. present a study on a novel CNN-based AI system for polyp characterization [6]. The authors showed that, by combining high-definition white-light and blue-light imaging (so-called multimodal imaging), the AI overall diagnostic accuracy was impressive (95.0 %), and higher than that of both experts (81.7 %) and novices (66.7 %). Interestingly, they also showed that the initial polyp characterization without AI (defined as “intuition”) had low diagnostic accuracy, which did not increase when using a dedicated classification (BLI Adenoma Serrated International Classification [BASIC] [3]), either for experts (intuition 79.5 % vs. BASIC 81.7 %; P = 0.14) or for novices (intuition 66.7 % vs. BASIC 66.5 %; P = 0.95). These findings emphasize the value of the AI system in the characterization process.

Nevertheless, despite its strengths, the paper has some methodological limitations that we need to highlight in order to put these results into perspective. First, the study involved only still, selected, high-quality images. Studies based on real-life videos are needed before a true estimate of AI performance can be reliably made. AI performance should be tested during real-life endoscopic procedures, as AI polyp characterization is affected by camera movements, peristaltic waves, bubbles/residues, and optical disturbances pertaining to light reflection and focus. As reported in the study by Mori et al. [7], within a clinical framework and according to the Preservation and Incorporation of Valuable endoscopic Innovations (PIVI) statements [1], the main outcome of any AI validation study should be clinically oriented. For this purpose, the study by van der Zander et al. [6] appears largely underpowered. Second, a tool for everyday clinical practice should demonstrate good discriminatory performance for every polyp histology. In the van der Zander et al. study, as in other studies, both the training and testing datasets had few sessile serrated lesions (SSL) or polyps with superficial invasion. Furthermore, the authors included SSLs within the neoplastic group, while some other AI characterization systems substantially equate SSLs to hyperplastic polyps or do not include them in the final analysis. Finally, it is well known that extensive training is needed for endoscopists to achieve acceptable results in polyp characterization, and the skill requires regular updates and, occasionally, retraining.

It has been argued that AI technology has the ability to decrease unwarranted variations and to improve the quality standards of all endoscopists. However, when dealing with AI systems, the relevance of the endoscopist’s background (defined as “intuition” by van der Zander et al. [6]) and the required training before AI use are likely to represent key factors in need of careful evaluation and definite standardization. Furthermore, there are concerns that nonexperts might passively accept the AI prediction without challenging it, thus reducing the endoscopist to a mere technician. However, AI is not foolproof and mistakes can occur. Therefore, only well-trained endoscopists with a strong basis in optical diagnosis will be able to accept or refuse the AI characterization output and give the final diagnosis. Thus, instead of abolishing the need for training and education, a high dedicated competence in polyp characterization without AI assistance becomes a prerequisite for anyone wishing to use AI in clinical practice. This is also relevant because legal issues of accountability of the final medical diagnosis and/or treatment can be raised.

In conclusion, polyp characterization remains a complex process and nowadays the indiscriminate application of AI, without addressing existing arguments, might lead to scientific, clinical, and legal issues. However, the preliminary results appear encouraging; some reports have demonstrated that AI-based characterization is cost effective, computerized systems based on training sets are easily upgradable, and the research will expand further. Once existing unresolved issues have been appropriately addressed, AI will be ready for prime time and AI-assisted polyp characterization will become current practice in every endoscopy facility. We are just at the beginning of an exciting, but long and winding road.

Publication History

Publication Date:
13 July 2021 (online)

© 2021. Thieme. All rights reserved.

Georg Thieme Verlag KG
Rüdigerstraße 14, 70469 Stuttgart, Germany