Automated Diagnosis of Various Gastrointestinal Lesions Using a Deep Learning-Based Classification and Retrieval...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Early diagnosis of gastrointestinal diseases is crucial for effective treatment and reducing life-threatening risks.
• Small gastrointestinal lesions are often undetectable during early examinations.
• Existing deep learning tools primarily focus on limited diseases (polyps, tumors) in specific GI tract areas.

Purpose of the Study:

• To develop a comprehensive computer-aided diagnosis (CAD) tool for diverse gastrointestinal diseases.
• To assist medical experts in improving the accuracy and efficiency of gastrointestinal disease diagnosis.

Main Methods:

• A framework combining a deep learning classification network with a retrieval method.
• The classification network predicts disease type from endoscopic images.
• A retrieval system presents relevant past cases to aid subjective validation by medical experts.

Main Results:

• Experiments conducted on 2 endoscopic datasets (52,471 frames, 37 classes).
• Achieved optimal performance metrics: 96.19% accuracy, 96.99% F1 score, 98.18% mean average precision, 95.86% mean average recall.
• The proposed framework significantly outperformed state-of-the-art methods.

Conclusions:

• The study presents a comprehensive CAD framework for identifying various gastrointestinal diseases.
• The method demonstrates superiority over recent approaches, showing potential for clinical application.
• The network's applicability extends to other medical imaging domains like CT, MRI, and ultrasound.