Constructing an artificial intelligence-assisted system for the assessment of gastroesophageal valve function based on the hill classification (with video).

Abstract

Objective: In the functional assessment of the esophagogastric junction (EGJ), the endoscopic Hill classification plays a pivotal role in classifying the morphology of the gastroesophageal flap valve (GEFV). This study aims to develop an artificial intelligence model for Hill classification to assist endoscopists in diagnosis, covering the entire process from model development, testing, interpretability analysis, to multi-terminal deployment.

Method: The study collected four datasets, comprising a total of 1143 GEFV images and 17 gastroscopic videos, covering Hill grades I, II, III, and IV. The images were preprocessed and enhanced, followed by transfer learning using a pretrained model based on CNN and Transformer architectures. The model training utilized a cross-entropy loss function, combined with the Adam optimizer, and implemented a learning rate scheduling strategy. When assessing model performance, metrics such as accuracy, precision, recall, and F1 score were considered, and the diagnostic accuracy of the AI model was compared with that of endoscopists using McNemar's test, with a p-value < 0.05 indicating statistical significance. To enhance model transparency, various interpretability analysis techniques were used, including t-SNE, Grad-CAM, and SHAP. Finally, the model was converted into ONNX format and deployed on multiple device terminals.

Results: Compared through performance metrics, the EfficientNet-Hill model surpassed other CNN and Transformer models, achieving an accuracy of 83.32% on the external test set, slightly lower than senior endoscopists (86.51%) but higher than junior endoscopists (75.82%). McNemar's test showed a significant difference in classification performance between the model and junior endoscopists (p < 0.05), but no significant difference between the model and senior endoscopists (p ≥ 0.05). Additionally, the model reached precision, recall, and F1 scores of 84.81%, 83.32%, and 83.95%, respectively. Despite its overall excellent performance, there were still misclassifications. Through interpretability analysis, key areas of model decision-making and reasons for misclassification were identified. Finally, the model achieved real-time automatic Hill classification at over 50fps on multiple platforms.

Conclusion: By employing deep learning to construct the EfficientNet-Hill AI model, automated Hill classification of GEFV morphology was achieved, aiding endoscopists in improving diagnostic efficiency and accuracy in endoscopic grading, and facilitating the integration of Hill classification into routine endoscopic reports and GERD assessments.