Constructing and validating vision transformer-based assisted detection models for atrophic gastritis: A...

Area of Science:

• Artificial Intelligence in Medicine
• Gastroenterology
• Medical Imaging Analysis

Background:

• Chronic atrophic gastritis (CAG) diagnosis relies on endoscopy, with lesion detection varying among endoscopists.
• Accurate identification and localization of atrophic lesions are crucial for effective CAG management.
• Vision transformer models offer potential for enhancing diagnostic accuracy in endoscopic procedures.

Purpose of the Study:

• To train and validate vision transformer-based models for detecting and localizing atrophic lesions in chronic atrophic gastritis (CAG).
• To assist endoscopists in improving the accuracy and consistency of CAG diagnosis.
• To identify both CAG and specific atrophic regions within the stomach using AI.

Main Methods:

• Retrospective collection of gastroscopy images (June 2019 - March 2023).
• Manual classification and annotation of images into CAG and chronic nonatrophic gastritis (CNAG) using Labelme software.
• Training of Swin transformer and SSFormer models on annotated images for detecting anatomical structures, CAG, and atrophic lesions.

Main Results:

• Swin transformer achieved 0.98 accuracy in recognizing gastric anatomical structures.
• Swin transformer demonstrated 0.91 accuracy, 0.95 specificity, and 0.86 sensitivity for CAG vs. CNAG detection, outperforming junior endoscopists (p < .05).
• SSFormer model showed lesion segmentation overlap exceeding 0.90, comparable to senior endoscopists.

Conclusions:

• Vision transformer models can effectively identify CAG, intragastric structures, and the extent of atrophy.
• These AI models show promise in increasing the accuracy of CAG diagnosis.
• The developed models can serve as valuable tools to support endoscopists in clinical practice.