Deep learning to predict esophageal variceal bleeding based on endoscopic images

Area of Science:

• Hepatology
• Medical Imaging
• Artificial Intelligence

Background:

• Esophageal varix (EV) bleeding is a severe complication of cirrhosis.
• Current prediction methods rely on experienced endoscopists but are often unreliable and inefficient.

Purpose of the Study:

• To evaluate the feasibility of using deep learning (DL) models to predict the 12-month risk of EV bleeding using endoscopic images.
• To compare the diagnostic performance of DL models against human endoscopists.

Main Methods:

• Six DL models were trained for binary classification of endoscopic images.
• Models were validated on an external dataset and compared with classifications by two endoscopists.

Main Results:

• EfficientNet achieved the highest accuracy (0.893) in the test dataset, outperforming endoscopists (0.800 and 0.763).
• AI assistance improved endoscopist accuracy by 17.3% and 19.0%.
• Model architecture influenced statistical agreement.

Conclusions:

• Deep learning models show feasibility in predicting 12-month EV bleeding risk from endoscopic images.
• AI-aided diagnosis offers a promising tool for improved cirrhosis management.