TL;DR
Artificial intelligence (AI) is rapidly advancing in digestive endoscopy, offering significant potential to enhance training by improving skill acquisition and procedural quality. However, widespread integration faces challenges related to ethical considerations, potential deskilling, and the need for more robust evidence on cost-effectiveness and objective educational outcomes.
What Problem Does This Paper Address?
The integration of Artificial Intelligence (AI) into medical training, particularly in specialized fields like digestive endoscopy, is a rapidly evolving area. While AI tools are increasingly being developed and applied for diagnostic and therapeutic purposes in endoscopy, their specific role and impact on the training of future endoscopists remain largely unexplored. This paper addresses this gap by providing a comprehensive narrative review of the current evidence on AI-assisted endoscopy training, identifying its potential benefits, exploring existing drawbacks, and envisioning future directions for its implementation. The authors aim to synthesize the existing literature to offer a clear picture of where AI stands in endoscopy education and what steps are needed for its effective and responsible adoption.
Methods
This study employed a systematic narrative review methodology. The researchers conducted a comprehensive search of the MEDLINE database, including articles published from the inception of the database up to January 31, 2025. The search strategy combined keywords related to AI (e.g., "artificial intelligence," "deep learning," "machine learning," "computer-aided") with terms for training ("training," "education") and various types of endoscopy ("luminal endoscopy," "capsule endoscopy," "hepatobiliary endoscopy," "therapeutic endoscopy"). Articles were included if they featured an educational or training component and explicitly explored AI applications in GI endoscopy training. Excluded were reviews, letters, editorials, comments, studies focused solely on model development, those lacking a training component, or those limited to simple performance comparisons between endoscopists and AI systems. A PRISMA flow diagram (Figure 1) illustrates the study selection process, detailing the number of records identified, screened, and ultimately included in the review.
Datasets
Not discussed in the paper. The paper focuses on a review of existing studies, rather than introducing new datasets or benchmarks.
Key Insights and Ideas
- AI is not just for diagnosis; it's a training tool: The paper highlights that AI's utility extends beyond simply detecting lesions. It can actively contribute to skill acquisition, procedural guidance, and quality assessment for trainees.
- Early career stage is where AI can make the biggest impact: AI-assisted tools show particular promise in helping novice endoscopists overcome the steep learning curve, potentially accelerating their proficiency and reducing errors.
- Beyond CADe and CADx: While computer-aided detection (CADe) and diagnosis (CADx) are well-established, the future of AI in training lies in more sophisticated applications like automated quality assessment, real-time procedural feedback, and personalized learning pathways.
- The "black box" problem is a major hurdle: The lack of transparency in how some AI algorithms arrive at their conclusions poses a significant challenge for ethical adoption and for building trust among trainees and educators.
- Deskilling is a real concern: Over-reliance on AI could lead to endoscopists losing fundamental skills, making them less capable when AI is unavailable or malfunctions. This necessitates a balanced approach that integrates AI as a supplement, not a replacement, for traditional training.
Key Results
The review identified 27 studies (11 RCTs and 16 prospective studies) that met the inclusion criteria. These studies demonstrated AI's potential across various endoscopic domains:
- Luminal Endoscopy: AI-assisted colonoscopy significantly improved adenoma detection rates (ADR) in trainees, with some studies showing increases of 13% absolute and 41% relative. CADx systems showed mixed results for trainees, with some studies indicating improved accuracy and others showing no significant benefit.
- Hepatobiliary Endoscopy: AI systems showed promise in improving structure/station recognition in endoscopic ultrasound (EUS) training, boosting trainees' accuracy from 67.2% to 78.4%. AI also demonstrated potential in lesion diagnosis and quality control for EUS and ERCP.
- Capsule Endoscopy: AI training systems reduced reading time and improved lesion detection in trainees. One study showed trainees' lesion recognition scores improved significantly after using a computer-based CE training system.
- Advanced Therapeutic Endoscopy: AI tools are being developed for real-time procedural phase recognition (e.g., for ESD and POEM), aiming to improve accuracy and provide feedback. AI also shows potential in outlining critical structures and detecting complications.
Limitations and Future Work
The authors acknowledge several limitations and suggest future directions:
- Limited Real-World Evidence: While promising, much of the existing evidence is based on pre-clinical studies or simulations. More clinical trials are needed to prove the real-world benefits of AI in endoscopy training.
- Cost-Effectiveness: The financial implications of integrating AI into training programs are still uncertain and require further investigation.
- Objective Educational Outcomes: More research is needed to assess the impact of AI on objective educational outcomes and long-term skill retention.
- Ethical and Practical Hurdles: Concerns regarding deskilling, over-reliance, automation bias, liability, informed consent, algorithmic bias, and equitable access across different training centers need to be addressed through structured implementation, quality assurance, and regulatory frameworks.
- Personalized Training: Future AI systems should evolve beyond task-specific applications to offer comprehensive and personalized training solutions, including continuous performance tracking and tailored feedback.
- Train-the-Trainer Integration: AI can also be used to enhance the skills of endoscopy trainers, providing feedback on their teaching methods and helping them upskill.
- Hybrid Approach: The authors advocate for a hybrid training approach that combines AI with traditional mentorship from experienced clinicians to leverage the strengths of both.
Summary updated