A Benchmark Dataset of Endoscopic Images and Novel Deep Learning Method to Detect Intestinal Metaplasia and Gastritis...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Endoscopic diagnosis of stomach diseases like intestinal metaplasia (IM) and gastritis atrophy (GA) relies heavily on radiologist expertise, leading to diagnostic inconsistencies and inefficiency.
• Current diagnostic methods face challenges due to inter-observer variability and the time-intensive nature of manual image analysis.

Purpose of the Study:

• To develop and validate a novel machine learning (ML) framework to enhance the accuracy and efficiency of endoscopic diagnosis for IM and GA.
• To address the limitations of subjective interpretation and time constraints in traditional endoscopic examinations.

Main Methods:

• A large dataset of 21,420 endoscopy images from White Light Imaging (WLI) and Linked Color Imaging (LCI) was curated and annotated.
• A novel ML model, 'local attention grouping,' inspired by the human visual system, was developed to extract key visual features without reducing image resolution.
• Ensemble learning and a dual transfer learning strategy were employed to improve feature extraction and model performance by leveraging both WLI and LCI data.

Main Results:

• The ML model achieved high diagnostic performance: 99.18% accuracy for IM and 97.12% for GA.
• Specific performance metrics included high sensitivity and specificity for both conditions, outperforming existing deep learning models.
• The method demonstrated superior accuracy, specificity, and sensitivity compared to current mainstream deep learning approaches.

Conclusions:

• The proposed ML framework significantly improves the accuracy and efficiency of diagnosing intestinal metaplasia and gastritis atrophy via endoscopy.
• This approach offers a reliable, automated solution to reduce diagnostic variability and enhance the clinical utility of endoscopic imaging.
• The developed model represents a state-of-the-art advancement in AI-driven gastrointestinal disease diagnosis.