Enhanced accuracy for classification of video capsule endoscopy images using multiple deep learning convolutional...

Area of Science:

• Gastroenterology
• Medical Imaging
• Artificial Intelligence

Background:

• Video capsule endoscopy (VCE) aids small intestine abnormality detection but struggles with massive image volumes.
• Current AI classification of VCE images has not reached clinical-grade diagnostic accuracy (>99%).

Purpose of the Study:

• To develop a highly accurate AI system for classifying various categories of unbounded VCE images.
• To overcome limitations in VCE image analysis using a novel transfer learning approach.

Main Methods:

• A transfer learning approach utilizing multiple convolutional neural networks (CNNs) was employed.
• The system automatically extracts features without requiring image segmentation, fine-tuning existing models for specific classifiers.
• 17 CNNs were combined to build, test, and validate AI models on over 16,000 VCE GI images.

Main Results:

• The combined 17-CNN deep learning approach achieved an overall diagnostic accuracy of 99.79%.
• Specific conditions like bleeding and foreign bodies were identified with 100% accuracy.
• High performance was validated through confusion matrices, precision, recall, and F1 scores.

Conclusions:

• Accurate AI deep learning models for unbounded VCE image classification have been developed.
• The system demonstrates potential for improved diagnosis of various medical conditions in clinical practice.