Visual Features for Improving Endoscopic Bleeding Detection Using Convolutional Neural Networks.

Area of Science:

• Medical imaging
• Computer vision
• Gastroenterology

Background:

• Endoscopic bleeding detection is crucial for patient care.
• Current methods may lack accuracy in identifying subtle visual cues.
• Automated detection systems can aid clinicians in real-time diagnosis.

Purpose of the Study:

• To develop and evaluate a novel method for endoscopic bleeding detection using image processing and deep learning.
• To incorporate domain-specific visual features into convolutional neural network (CNN) models.
• To compare the performance of feature-enhanced CNNs against baseline models.

Main Methods:

• Defined high-level visual features of endoscopic bleeding incorporating domain knowledge.
• Developed feature descriptors to automatically extract these features, generating grayscale activation maps.
• Integrated feature maps with original color channels as input for CNNs (Resnet, VGG).
• Conducted comparative experimental evaluation using Receiver Operating Characteristic Area Under the Curve (ROC AUC).

Main Results:

• Feature-extended CNN models demonstrated improved classification performance compared to baseline models.
• The proposed method showed significant advantages for both Resnet and VGG architectures.
• Feature engineering enhanced the ability of CNNs to detect endoscopic bleeding.

Conclusions:

• Incorporating domain-specific visual features significantly boosts the performance of CNNs for endoscopic bleeding detection.
• The proposed feature-extension approach offers a promising direction for improving automated diagnostic tools in endoscopy.
• This method has the potential to enhance diagnostic accuracy and patient outcomes in endoscopic procedures.