A Deep Learning Application of Capsule Endoscopic Gastric Structure Recognition Based on a Transformer Model

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Accurate diagnosis of gastric lesions in capsule endoscopy relies on effective gastric structure recognition systems.
• Deep learning, particularly transformer models utilizing self-attention, shows promise for gastrointestinal image analysis.
• Current deep learning applications in endoscopic image recognition require further development for clinical utility.

Purpose of the Study:

• To develop an artificial intelligence (AI) model for identifying gastric structures in capsule endoscopy images.
• To enhance the clinical applicability of deep learning techniques in endoscopic image recognition.
• To evaluate the diagnostic performance of the AI model in comparison to human endoscopists.

Main Methods:

• Utilized 3343 wireless capsule endoscopy videos for unsupervised pretraining, with 2433 for training and 118 for validation.
• Selected fifteen upper gastrointestinal (GI) structures for quantitative examination quality assessment.
• Compared the AI model's classification performance against endoscopists using accuracy, sensitivity, specificity, and predictive values.

Main Results:

• The transformer-based AI model demonstrated high diagnostic accuracy in gastric structure recognition.
• The AI model achieved a macroaverage accuracy of 99.6%, sensitivity of 96.4%, and specificity of 99.8% for identifying 15 upper GI structures.
• The AI model exhibited a high level of interobserver agreement with experienced endoscopists.

Conclusions:

• The transformer-based AI model accurately evaluates gastric structure information from capsule endoscopy.
• The AI model performs comparably to endoscopists in evaluating gastric structures.
• This AI tool can significantly assist physicians in diagnosing from extensive image datasets, thereby improving examination efficiency.