Recognizing IgA-class endomysial antibody equivalent binding patterns on monkey liver substrate through EfficientNet...

Area of Science:

• Medical Diagnostics
• Artificial Intelligence in Medicine
• Immunology

Background:

• Immunoglobulin A (IgA) endomysial antibody (EMA) testing is crucial for diagnosing celiac disease.
• Current EMA test interpretation is subjective and labor-intensive, requiring expert human analysis.
• Automating EMA test interpretation using deep learning could enhance diagnostic efficiency and objectivity.

Purpose of the Study:

• To evaluate the performance of EfficientNet and EfficientNetV2 deep learning architectures for automated IgA EMA test interpretation.
• To assess classification accuracy across binary, three-class, and four-class scenarios, including gray zone cases.
• To explore the potential of explainable AI (HiRes-CAM) in understanding deep learning model decisions for EMA-eq tests.

Main Methods:

• Utilized EfficientNet and EfficientNetV2 architectures for image classification of IgA EMA equivalent (EMA-eq) tests.
• Trained and tested models on 368 clinical samples across different classification tasks (binary, three-class, four-class).
• Employed HiRes-CAM for visualizing and explaining the deep learning model's interpretation process.

Main Results:

• EfficientNetV2-S achieved high accuracy: 99.37% (binary), 95.28% (three-class), and 86.98% (four-class).
• Medium-sized deep learning architectures outperformed larger ones, contrary to expectations.
• Higher input resolution (640x640) and architectural innovations in EfficientNet-V2 contributed to superior performance.

Conclusions:

• Deep learning models can achieve expert-level performance in interpreting IgA EMA-eq tests.
• Automated interpretation offers a more standardized, efficient, and objective approach to celiac disease diagnosis.
• This technology has the potential to reduce the workload on specialist medical staff.