An End-to-End Deep Learning System for Gastrointestinal Bleeding Detection and Quantification in Wireless Capsule...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Gastrointestinal bleeding detection via wireless capsule endoscopy (WCE) is crucial but labor-intensive.
• Manual review of WCE images is time-consuming and prone to missed diagnoses.

Purpose of the Study:

• To develop and evaluate a deep learning framework for automated bleeding detection, localization, and quantification in WCE.
• To reduce the diagnostic burden and improve accuracy in WCE analysis.

Main Methods:

• Integrated three deep learning models: 2D-CNN for frame classification, 3D-CNN for temporal analysis, and U-Net for pixel-level segmentation.
• Trained and validated models using expert-annotated WCE datasets with pixel-level ground truth.

Main Results:

• 2D-CNN and 3D-CNN achieved high classification performance (AUCs 0.9986 and 0.9971).
• U-Net demonstrated strong segmentation accuracy (Dice 0.93, IoU 0.8677, overall 97.25%).
• The integrated framework surpassed previous methods in bleeding detection, localization, and quantification.

Conclusions:

• The end-to-end deep learning framework provides accurate automated bleeding detection and severity assessment in WCE.
• Potential to enhance clinical decision-making and streamline gastrointestinal diagnostic workflows.