Endoscopic Diagnosis of Eosinophilic Esophagitis Using a Multi-Task U-Net: A Pilot Study

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Eosinophilic esophagitis (EoE) diagnosis is challenging due to subtle endoscopic findings.
• Accurate identification of EoE is crucial for timely intervention and management.

Purpose of the Study:

• To develop and validate a robust convolutional neural network (CNN) for identifying and classifying EoE in endoscopic images.
• To compare the CNN's performance against established models and human endoscopists.

Main Methods:

• A multi-task U-Net with auxiliary classifiers (scaU-Net) was trained on 548 EoE and 297 normal endoscopic images.
• Datasets were labeled based on EREFS features: edema, rings, exudates, and furrows.
• Performance was evaluated using accuracy, AUROC, Grad-CAM, and comparison with 25 endoscopists.

Main Results:

• The proposed sca4U-Net achieved high performance: 86.9% accuracy and 0.93 AUROC.
• Deepest skip connections in the U-Net architecture were critical for optimal results.
• The AI model demonstrated superior performance compared to endoscopists of varying experience levels.

Conclusions:

• The developed AI model shows robust performance in detecting EoE, comparable to expert endoscopists.
• This AI tool can assist endoscopists at all levels in the early diagnosis of EoE.
• The scaU-Net model offers a promising approach for improving EoE detection rates.