Surfing the AI wave: Insights and challenges

The interest in artificial intelligence (AI) and its applications in gastrointestinal endoscopy remains undiminished [1]. The huge mass of studies being published on the topic shows that researchers and practitioners are keen to produce innovations, publish new insights, and, at the same time, keep abreast of current trends and developments. The AI wave does not seem to have broken yet, but as we ride the wave, critical thinking and, most importantly, reappraisal of current data and approved products is paramount [2].

Troya et al. may have had this in mind when they designed their present study [3]. The authors compared the performance of various computer-aided detection (CADe) systems for identification of polyps using the same benchmark dataset. For their comparisons, Troya et al. included the earlier and current versions of the GI Genius CADe system, two different detection modes of the EndoAID CADe system (Type A and B), and a freely available open-access system called EndoMind. The dataset comprised 101 colonoscopy videos; 45 had at least one polyp.

“…human-AI interactions will be strongly influenced by the experience and receptiveness of the physician-user…”

The primary outcome, the per-polyp sensitivity, was comparable among all systems. In addition to this, the authors tried to improve the granularity of their sensitivity comparisons by calculating the per-frame sensitivity for each polyp: the number of frames with a correctly identified image of a particular polyp divided by the number of benchmark frames with an image of that polyp. For the per-frame sensitivities, a difference was seen across the various comparisons, including mean per-frame sensitivities of 50.63% and 67.85%, respectively, for the earlier and the later version of GI-Genius, 65.60% and 52.95%, respectively, for EndoAID Types A and B, and 60.22% for the freely available EndoMind. Other comparisons included the false-positive rate, in which the Type B EndoAID version scored best. For the time-to-first-detection, the more sensitive Type A EndoAID and version 2 of GI Genius performed best.

To better understand the significance of Troya’s results, one needs to take a step back and see the overall picture that the authors have tried to paint. For a standard practitioner, the mass of information, innovations, and products flooding the scene is difficult to comprehend, differentiate and categorize. Troya et al. have tried to compare the most popular CADe systems. Even though these results were produced “only” on a video dataset and not in real time, the authors should be commended for providing this pioneering information. A closer look at the immense number of frames included in the final assessment offers better insights into the validity of the presented data.

This leaves us with the decisive question: Quo vadis artificial intelligence? [4]. The current work and its motivation provide at least part of the answer. Constant interrogations, re-evaluations, and critical reassessments must accompany all our scientific endeavors around the “buzzword” topic of AI. New technology should not be adopted blindly but followed up with well-planned and executed trials. In addition, future AI research should always include the more unpredictable “human factor.” The implementation of AI tools by practicing physicians will uncover an additional aspect that must be considered because human–AI interactions will always produce different results from the “naked” standalone performance of AI systems [5] [6]. More importantly, human–AI interactions will be strongly influenced by the experience and receptiveness of the physician-user [7].

Ocean waves can become even more enticing to ride as they roll towards the shore. They are strong, high, and can change their form rapidly. Similarly, as we all ride slowly ashore on the AI wave, let us always pair critical reappraisal with acceptance and open-mindedness, to the ultimate benefit of our patients.

Publication History

Article published online:
27 October 2023

• References

• 1 Rey JF. Artificial intelligence in digestive endoscopy: recent advances. Curr Opin Gastroenterol 2023; 39: 397-402 DOI: 10.1097/mog.0000000000000957. (PMID: 37522929)
• 2 Frake CO. Cognitive maps of time and tide among medieval seafarers. Man New Series 1985; 20: 254-270 DOI: 10.2307/2802384.
• 3 Troya J, Sudarevic B, Krenzer A. et al. Direct comparison of multiple computer-aided polyp detection systems. Endoscopy 2023; 55 DOI: 10.1055/a-2147-0571.
• 4 Jiang Y, Li X, Luo H. et al. Quo vadis artificial intelligence?. Discov Artif Intell 2022; DOI: 10.1007/s44163-022-00022-8.
• 5 Reverberi C, Rigon T, Solari A. et al. Experimental evidence of effective human-AI collaboration in medical decision-making. Sci Rep 2022; 12: 14952 DOI: 10.1038/s41598-022-18751-2. (PMID: 36056152)
• 6 Scheppach MW, Rauber D, Stallhofer J. et al. Detection of duodenal villous atrophy on endoscopic images using a deep learning algorithm. Gastrointest Endosc 2023; 97: 911-916 DOI: 10.1016/j.gie.2023.01.006. (PMID: 36646146)
• 7 Messmann H, Ebigbo A, Hassan C. et al. How to integrate artificial intelligence in gastrointestinal practice. Gastroenterology 2022; 162: 1583-1586 DOI: 10.1053/j.gastro.2022.02.029. (PMID: 35196540)