Deep learning-based anatomical position recognition for gastroscopic examination.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Gastroenterology

Background:

• Gastroscopic examination is crucial for detecting upper gastrointestinal lesions.
• Current gastroscopy training faces challenges due to strict image archiving requirements for junior doctors.
• Accurate anatomical position identification is essential for comprehensive examination and documentation.

Purpose of the Study:

• To develop an automated deep learning-based system for recognizing anatomical positions during gastroscopic examinations.
• To enhance the training and diagnostic capabilities of medical professionals in endoscopy.

Main Methods:

• A dataset of 17,182 gastroscopic images across eight anatomical categories was utilized.
• The MogaNet convolutional neural network model was employed for anatomical position identification.
• Model performance was rigorously evaluated using sensitivity, precision, and F1-score metrics.

Main Results:

• The proposed MogaNet model achieved superior performance compared to ResNet, GoogleNet, and SqueezeNet.
• Average sensitivity, precision, and F1-score for the MogaNet model were 0.963, 0.964, and 0.964, respectively.
• Statistical analysis confirmed a significant improvement (p < 0.05) over existing models.

Conclusions:

• The developed deep learning method demonstrates excellent performance in recognizing anatomical positions during gastroscopy.
• This technology can assist junior doctors in meeting examination completeness and image archiving standards efficiently.
• Automated position recognition holds promise for improving the quality and consistency of endoscopic training and practice.