Expert-AI Collaborative Training for Novice Endoscopists: A Path to Enhanced Efficiency

Zhen Zhang, B. Chen, Ling Du, Quan‐Lin Li, Yan Zhu, Pei‐Yao Fu, Wen‐Zheng Qin, Huankai Shou, Pingting Gao, Xin‐Yang Liu, Meng‐Jiang He, Zihan Geng, Shuo Wang, Ping‐Hong ZhouBioengineering 2025 · DOI

TL;DR

This paper introduces EndoAdd, an AI-powered system designed to train novice endoscopists by providing real-time feedback on blind spots and photodocumentation. The study demonstrates that EndoAdd significantly reduces blind spots and improves lesion detection compared to traditional training methods, accelerating the learning curve and enhancing procedural quality.

What Problem Does This Paper Address?

Gastrointestinal endoscopy, particularly esophagogastroduodenoscopy (EGD), is a crucial diagnostic tool, but training novice endoscopists is challenging. Traditional methods often suffer from inefficiency, inconsistency, and a shortage of expert supervisors. Novices struggle with recognizing anatomical landmarks and ensuring complete mucosal inspection, leading to a higher risk of missed lesions and variable training outcomes. Existing AI tools for endoscopy primarily focus on detection or diagnosis, with a lack of systems specifically designed for structured, real-time training support for beginners. This gap highlights the need for innovative solutions to improve the quality and efficiency of endoscopy education.

Methods

The core of the EndoAdd system is a deep learning model based on the XceptionNet architecture, a convolutional neural network known for its efficiency in image recognition. This model was trained to classify EGD images into 26 predefined anatomical categories, enabling real-time identification of the visualized region. The system then uses this classification to generate a dynamic "blind spot map" and provide immediate feedback to the trainee.

Key methodological aspects include:

  • Real-time Feedback: The system provides live guidance, highlighting areas that have not been adequately visualized.
  • Photodocumentation Assistance: It also helps ensure comprehensive documentation by tracking visualized areas.
  • Randomized Controlled Trial (RCT): The study employed an RCT design to compare the EndoAdd group with a traditional training group.
  • Phased Training: The study involved a training phase, a practicing phase (with and without AI assistance), and a testing phase.
  • Data Augmentation: Techniques like random rotations, flipping, cropping, and color jitter were used to enhance model generalizability.
  • Hyperparameter Optimization: Grid search was employed to fine-tune learning rates, batch sizes, and dropout rates.
  • Model Validation: Stratified five-fold cross-validation and performance metrics like accuracy, sensitivity, specificity, and AUC were used.

Datasets

The study utilized a custom dataset comprising EGD images from 5000 patients. For model development and validation, images from 500 patients were selected and annotated. This dataset was divided into:

  • Training set: 80% of patients (400 patients, 35,974 images)
  • Validation set: 10% of patients (50 patients, 5415 images)
  • Test set: 10% of patients (50 patients, 5564 images)

The images were annotated into 27 categories, including 26 upper gastrointestinal tract sites and a "NA" (not applicable) class. The annotation process involved junior and senior endoscopists to ensure quality control.

Key Insights and Ideas

  • AI as a "Virtual Trainer": The study demonstrates that AI can effectively act as a real-time training director, providing consistent and objective feedback that complements human supervision.
  • Accelerated Learning Curve: EndoAdd significantly reduced the number of training cases required for novices to achieve proficiency, suggesting a faster and more efficient learning process.
  • Bridging the Gap: The research contributes to bridging the gap between AI algorithm development and practical, standardized deployment in medical education.
  • Beyond Blind Spots: While blind spot reduction is a primary benefit, the AI system also showed improvements in photodocumentation completeness, a crucial aspect for accurate diagnosis and communication.
  • Potential for Standardization: AI-driven training offers a path towards standardizing endoscopy education, reducing variability in skill acquisition.

Key Results

The EndoAdd system achieved impressive performance:

  • Image Classification Accuracy: 98.0% overall accuracy and a mean AUC of 0.984 on the test set.
  • Reduced Blind Spots: The EndoAdd group had significantly fewer blind spots compared to the traditional training group (p < 0.01).
  • Improved Photodocumentation: Completeness of photodocumentation was significantly higher in the EndoAdd group (74% vs. 59%, p < 0.01).
  • Enhanced Lesion Detection: While not statistically significant across all lesions, the AI group showed trends towards improved detection in specific areas.
  • Reduced Examination Time: Examination time was reduced without compromising diagnostic accuracy.
  • Fewer Training Cases: EndoAdd trainees required a median of 28 cases compared to 41 for traditional trainees (p = 0.003).
  • Comparable Diagnostic Accuracy: Neither training method compromised diagnostic accuracy for common lesions.

Limitations and Future Work

The authors acknowledge several limitations:

  • Dataset Bias: The training dataset's single-center origin and East Asian predominance may limit external validity for ethnically diverse populations.
  • Ceiling Effect: The system's performance metrics approached ceiling levels for many lesion types, potentially masking the full incremental value of AI.
  • Limited Sample Size: The pilot study's small cohort restricts generalizability and the ability to capture inter-individual variability.
  • Blinding Challenges: Complete blinding of trainers during practice phases was unfeasible due to the visible AI system.
  • Lack of Longitudinal Follow-up: The study did not assess long-term skill retention or autonomous application post-AI training.
  • Cost-Effectiveness: A formal cost-benefit analysis was not conducted, and upfront implementation costs may be a barrier.
  • Anecdotal Feedback: Qualitative feedback was informal and not obtained through structured questionnaires.

Future work should focus on:

  • Multicenter Trials: Validating the system's scalability and educational sustainability across diverse settings.
  • Longitudinal Studies: Assessing long-term skill retention and broader implementation strategies.
  • More Challenging Cases: Incorporating more subtle or rare lesion cases to address the ceiling effect.
  • Structured Qualitative Data: Employing standardized, quantitative measures to capture trainee perspectives.
  • Cost-Effectiveness Analysis: Investigating the economic benefits of AI-assisted training.
  • Enhanced AI Algorithms: Developing more sophisticated lesion characterization and real-time procedural guidance.
  • Addressing Ethical Considerations: Ensuring responsible adoption of AI, including patient consent and maintaining the human aspect of care.

Summary updated

Abstract

The AI-assisted EndoAdd system improved novice endoscopist performance, reducing blind spots and enhancing lesion detection. AI systems like EndoAdd show potential in accelerating endoscopy training and improving procedural quality.
BibTeX
@inproceedings{zhang2025expert,
  author = {Zhen Zhang and B. Chen and Ling Du and Quan‐Lin Li and Yan Zhu and Pei‐Yao Fu and Wen‐Zheng Qin and Huankai Shou and Pingting Gao and Xin‐Yang Liu and Meng‐Jiang He and Zihan Geng and Shuo Wang and Ping‐Hong Zhou},
  title = {{Expert-AI Collaborative Training for Novice Endoscopists: A Path to Enhanced Efficiency}},
  booktitle = {Bioengineering},
  year = {2025},
  doi = {10.3390/bioengineering12060582},
}
Figure