Deep learning site classification model for automated photodocumentation in upper GI endoscopy (with video).

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Photodocumentation in ERCP (esophagogastroduodenoscopy) can be automated using deep learning (DL) for anatomic site classification.
• ERCP video data often contain suboptimal frames (blurry, off-center) hindering computational analysis.
• A DL model was developed to extract high-quality frames from ERCP videos for improved anatomic classification.

Purpose of the Study:

• To develop and validate a DL model for extracting high-quality frames from ERCP video data.
• To classify extracted frames into one of eight predefined anatomic sites.
• To enable automated and standardized photodocumentation in ERCP procedures.

Main Methods:

• A DL algorithm with two image filters was used to identify high-quality frames (centered, clear).
• The model classified frames into eight anatomic sites: esophagus, gastroesophageal junction, stomach (body, fundus, angularis, antrum), and duodenum (bulb, distal).
• Model training and validation involved over 8,000 still images and 26,000 video-derived images, with external validation on over 2,000 images.

Main Results:

• Internal testing achieved 98.1% accuracy and F1 scores of 90.0%-99.0%.
• External validation demonstrated 95.0% accuracy and F1 scores of 92.0%-97.0%.
• Application to ERCP videos yielded accuracies between 89.7% and 94.8% across different frame window sizes.

Conclusions:

• A DL model effectively extracts high-quality frames from ERCP videos for anatomic classification.
• This approach facilitates automated photodocumentation, ensuring consistent study quality.
• The model aids in video indexing for efficient, annotated review of ERCP procedures.