Deep learning-based clinical decision support system for gastric neoplasms in real-time endoscopy: development and...

Area of Science:

• Gastroenterology
• Medical Imaging
• Artificial Intelligence

Background:

• Deep learning models have shown promise in predicting gastric lesion characteristics from endoscopic images.
• Previous models focused on histopathology and invasion depth prediction.
• Automated detection and classification of gastric neoplasms are crucial for timely diagnosis and treatment.

Purpose of the Study:

• To establish and validate a deep learning-based clinical decision support system (CDSS) for real-time automated detection and classification of gastric neoplasms.
• To assess the system's performance in diagnosing lesion type and predicting invasion depth during endoscopy.

Main Methods:

• Utilized 5017 endoscopic images for training the deep learning models.
• Validated the lesion detection model in a randomized pilot study of 2524 real-time procedures, comparing CDSS-assisted with conventional endoscopy.
• Validated the lesion classification model using a prospective multicenter external test with 3976 novel images.

Main Results:

• The lesion detection model achieved a 95.6% detection rate in internal testing.
• CDSS-assisted endoscopy showed a trend towards a higher lesion detection rate (2.0% vs. 1.3%) in the randomized study.
• The CDSS demonstrated 81.5% accuracy for four-class classification and 86.4% for binary invasion depth prediction in the external test.

Conclusions:

• The developed CDSS shows significant potential for real-world clinical application in endoscopy.
• The system exhibits high performance in detecting and classifying gastric lesions.
• This technology can aid clinicians in real-time diagnosis and management of gastric neoplasms.