Esophageal squamous cell carcinoma (ESCC) remains a significant global health burden, with persistently high incidence and mortality rates [1]. Early detection and timely treatment are essential for improving patients outcomes, with esophagogastroduodenoscopy playing an irreplaceable role in this process [2]. The endoscopic features of superficial ESCC are however often subtle compared with the surrounding mucosa, necessitating the integration of multiple modalities, such as white-light imaging (WLI), narrow-band imaging (NBI), and iodine staining, for effective screening. The detection rate of superficial ESCC varies considerably, influenced by disparities in endoscopic equipment, as well as the skills and experience of individual endoscopists.

In recent years, the widespread application of artificial intelligence (AI) in medicine has provided endoscopists with innovative tools to aid in diagnosis. Eisuke Nakao and colleagues have developed a series of AI-assisted diagnostic systems [3] [4] [5] that address multiple challenges in the diagnosis of ESCC. These systems encompass various applications, from detecting ESCC in static images to assessing the invasion depth of ESCC, and even recognizing ESCC in real-time videos. Their article published in this issue of Endoscopy presents the clinical application of these advancements, offering a comprehensive evaluation of their diagnostic performance and effectiveness [6].

The authors conducted a prospective, single-center, exploratory, and randomized controlled trial. The results of AI-assisted diagnosis for ESCC in this study were less promising than expected. Notably, the AI system was not integrated into the primary endoscopy monitor but was displayed on an external screen. During the procedure, endoscopists in the AI group had access to both the standard monitor and the AI screen, where potential ESCC lesions were marked for reference. In contrast, the control group performed routine endoscopic examinations without AI support. In both groups, endoscopists initially examined the esophagus using WLI, followed by NBI. After completing the examination of the upper gastrointestinal tract, the esophagus was further assessed using iodine staining. The findings showed no significant difference in the detection rates of ESCC between the AI and control groups. Subgroup analyses, including comparisons between expert and nonexpert endoscopists, similarly demonstrated no improvement with AI assistance.

We have recently conducted a multicenter, tandem, double-blind, randomized controlled trial involving 73 endoscopists from 12 hospitals and 11715 participants [7]. The results showed that the use of our AI system significantly improved the detection rate of superficial ESCC and precancerous lesions. Notably, our AI system was directly integrated into the endoscopy monitor, operating on a single screen to align with clinical workflows. One possible reason for the negative results observed in the study by Nakao and colleagues [6] is the lack of a real-time AI display on the same screen, which may have introduced distractions for the endoscopists, a concern the authors mentioned in their discussion. Additionally, a retrospective review of videos and static images from their study revealed that 20% of lesions were undetectable using both WLI and NBI and could only be identified through iodine staining. This may have further contributed to their negative findings. Furthermore, although the AI system flagged potential malignancies, endoscopists did not always perform biopsies based solely on the AI’s suggestions.

In summary, when using AI for diagnostic assistance, we must recognize two key points. First, AI serves only as an auxiliary tool; the final decision rests with the endoscopists. The extent to which AI alerts are accepted or disregarded depends on the endoscopists' knowledge structure, level of experience, and the specific circumstances during the diagnostic process. Second, there are limitations in training AI based on endoscopists’ recognition and judgment. If endoscopists cannot accurately identify certain conditions, such as lesions detectable only through iodine staining, teaching the AI to recognize these situations may prove challenging.

Despite the significant progress of AI in gastrointestinal endoscopy and its increasing sensitivity and specificity, it is undeniable that AI has not yet surpassed the performance of expert endoscopists. Nevertheless, we must recognize the promising potential and significant implications of AI in the auxiliary diagnosis of ESCC. By using AI appropriately, we can enhance our own diagnostic capabilities. We look forward to further advancements in AI algorithms, such as generative AI, which incorporate logical reasoning and emergent knowledge capabilities.

Article published online:
28 November 2024

© 2024. Thieme. All rights reserved.

Georg Thieme Verlag KG
Oswald-Hesse-Straße 50, 70469 Stuttgart, Germany