Deep Learning Methods for Anatomical Landmark Detection in Video Capsule Endoscopy Images

Area of Science:

• Gastroenterology
• Medical Imaging
• Artificial Intelligence

Background:

• Video capsule endoscopy (VCE) offers non-invasive visualization of the entire gastrointestinal (GI) tract.
• Traditional endoscopy is invasive and limited in reach, while VCE generates extensive image data.
• Analyzing hours of VCE footage is time-consuming and costly, hindering automated diagnosis.

Purpose of the Study:

• To develop automated methods for VCE image analysis.
• To accurately detect the anatomical location of the capsule within the GI tract from VCE images.
• To compare the performance of deep Convolutional Neural Network (CNN) models for VCE image analysis.

Main Methods:

• Four deep CNN models (VGG-Net, GoogLeNet, AlexNet, ResNet) were evaluated.
• Models were trained for feature extraction and anatomical part detection in VCE images.
• Performance was assessed using accuracy, precision, recall, and F1-score.

Main Results:

• VGG-Net demonstrated the highest average accuracy, precision, recall, and F1-score.
• VGG-Net outperformed GoogLeNet, AlexNet, and ResNet in capsule localization.
• The findings indicate VGG-Net's suitability for automated VCE image analysis.

Conclusions:

• Automated capsule localization in VCE is crucial for efficient GI disease diagnosis.
• Deep learning models, particularly VGG-Net, show significant promise for VCE image analysis.
• This approach can streamline the review process and improve diagnostic accuracy.