Deep Transfer Learning for Automated Intestinal Bleeding Detection in Capsule Endoscopy Imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Gastroenterology

Background:

• Capsule endoscopy (CE) is a vital tool for visualizing the small intestine.
• Manual analysis of CE images for abnormalities like bleeding is time-consuming and labor-intensive.
• Automated detection of bleeding in CE images is crucial for timely diagnosis and treatment.

Purpose of the Study:

• To develop a computer-aided diagnostic (CAD) tool for automated analysis of capsule endoscopic (CE) images.
• To accurately detect small intestinal abnormalities, specifically bleeding.
• To delineate bleeding zones within identified images.

Main Methods:

• A deep learning framework utilizing convolutional neural networks (CNNs) was developed.
• Transfer learning with a pre-trained AlexNet model was employed for bleeding identification.
• Semantic segmentation using a SegNet deep neural network was used for bleeding zone delineation.

Main Results:

• The framework achieved high F1 scores of 98.49% and 88.39% on two public datasets for bleeding detection.
• Bleeding zone identification yielded a 94.42% global accuracy and 90.69% weighted intersection over union (IoU).
• Performance demonstrated consistent advances compared to state-of-the-art methods.

Conclusions:

• The proposed deep learning framework significantly improves bleeding detection and delineation in CE images.
• The system offers substantial savings in annotation time and human labor compared to manual inspection.
• The automated approach enhances diagnostic accuracy and has the potential to reduce overall CE costs, increasing accessibility.