WEO Newsletter: The Impact of Artificial Intelligence on Gastrointestinal Endoscopy

Abstract

WEO Newsletter Editor: Nalini M Guda MD, MASGE, AGAF, FACG, FJGES

Michael B. Wallace MD MPH

John C. Anderson Professor of Medicine, Mayo Clinic Florida

Artificial intelligence (AI) is rapidly transforming the field of gastrointestinal (GI) endoscopy, enhancing diagnostic accuracy, efficiency, and workflow optimization. AI-driven technologies, including deep learning algorithms and computer-aided detection and diagnosis (CADe/CADx), are being increasingly integrated into endoscopic practice, particularly in areas such as adenoma detection, polyp classification, Barrett's esophagus evaluation, early gastric cancer detection, indeterminate bile duct strictures, and bowel preparation classification. This review explores these key areas and the implications of AI on endoscopic workflows.

Colorectal cancer (CRC) prevention relies heavily on the detection and removal of adenomatous colon polyps during colonoscopy. AI-powered CADe systems have been developed to enhance adenoma detection rates (ADR) by identifying subtle lesions that may be overlooked by endoscopists. Studies have shown that AI-assisted colonoscopy increases ADR, reduces polyp miss rates, and improves overall procedural quality. By providing real-time visual alerts and bounding boxes around suspected polyps, AI enables more effective and standardized detection.

Beyond detection, AI plays a critical role in polyp characterization (CADx), aiding endoscopists in distinguishing between neoplastic and non-neoplastic lesions. AI algorithms trained on large datasets of histologically confirmed polyps can provide real-time classification, potentially reducing the need for unnecessary polypectomies. Technologies such as narrow-band imaging (NBI) and confocal laser endomicroscopy, when combined with AI, can enhance the accuracy of in vivo histological assessments. While nearly all polyps in the proximal colon should be removed endoscopically, making pre resection diagnosis less relevant, there remains a value in the rectosigmoid where small hyperplastic polyps can be left in Situ when the endoscopist is confident that they are hyperplastic. Furthermore, large polyps must be stratified into those that are noninvasive, superficially invasive, and deeply invasive to guide therapies such as standard endoscopic mucosal resection, endoscopic submucosal dissection and surgery respectively.

AI has also been recently shown to facilitate polyps size classification which is notoriously variable and has impact on surveillance recommendations particularly for polyps 10 mm or larger. Having objective measures of size should further standardize surveillance recommendations.

Barrett's esophagus (BE) is a precursor to esophageal adenocarcinoma, requiring precise surveillance and risk stratification. AI-based systems have been developed to detect BE and its progression to dysplasia by analyzing endoscopic images with high sensitivity and specificity. These algorithms can aid in targeting biopsy locations and standardizing reporting, thereby reducing interobserver variability and improving early detection of high-risk lesions.

Early gastric cancer (EGC) can be challenging to detect due to its subtle and heterogeneous appearance. AI-powered image analysis has proven the ability to enhance EGC detection by recognizing patterns associated with malignancy that may not be readily apparent to the human eye. AI can aid in real-time decision-making, guiding endoscopists toward targeted biopsies and improving diagnostic yield.

The differentiation of benign and malignant bile duct strictures remains a clinical challenge. AI-driven tools using deep learning and radiomics analysis of endoscopic ultrasound (EUS) and cholangioscopy images have shown promise in improving diagnostic accuracy. By integrating AI into these modalities, clinicians may reduce the need for invasive procedures and hasten appropriate management strategies.

Adequate bowel preparation is essential for high-quality colonoscopy. AI has been employed to assess and grade bowel cleanliness in real time, providing feedback to the endoscopist regarding whether additional washing or suctioning is needed. Standardized AI-driven bowel preparation scoring could improve procedural efficiency and reduce the likelihood of incomplete examinations.

The integration of AI into GI endoscopy has the potential to significantly improve workflow by improving efficiency, reducing cognitive load, and enhancing decision-making. AI-driven tools can streamline pre-procedure planning by assessing patient risk profiles and improving scheduling. During procedures, real-time AI assistance can improve detection and characterization of lesions, leading to higher-quality examinations with less variability among endoscopists. Post-procedure, AI can aid in automated report generation, reducing documentation time and enhancing compliance with quality metrics. Recent advances in ambient listening technology combined with large language models have allowed significantly more efficient workflow in the clinical space and should be directly applicable to endoscopy. One caveat is that the data source for endoscopy is actually the video image which could be combined with voice annotation to generate procedure reports of high quality. Additionally, AI can provide personalized post-procedure recommendations and aid in training by offering real-time feedback to endoscopists. These advancements collectively contribute to a more efficient, accurate, and standardized approach to GI endoscopy, ultimately benefiting both clinicians and patients.

The successful implementation of AI in GI endoscopy depends on effective human-AI interaction. AI should be seen as an augmentation tool rather than a replacement for endoscopists. Clinicians must be trained to interpret AI-generated insights, integrating them into their clinical decision-making process while maintaining oversight. Additionally, trust in AI recommendations is crucial; endoscopists should validate AI outputs against clinical judgment and histopathological findings. Understanding AI's limitations, including potential biases and false positives, will be essential in ensuring its proper use in practice.

AI has the potential to revolutionize training in GI endoscopy by providing objective, real-time feedback to trainees. AI-assisted simulation platforms can offer personalized learning experiences, helping young physicians develop technical skills in a controlled environment. Additionally, AI can track performance metrics such as adenoma detection rate and withdrawal time, allowing educators to provide data-driven mentorship. By integrating AI into training programs, medical institutions can standardize education, enhance competency assessment, and ultimately improve the quality of endoscopic care.

As AI technology continues to evolve, its integration into GI endoscopy will play a crucial role in enhancing patient outcomes, reducing variability in diagnosis, and improving procedural workflows. Future research and regulatory considerations will be essential in ensuring the safe and effective implementation of AI-driven endoscopic tools in clinical practice.