Video Capsule Endoscopy Classification using Focal Modulation Guided Convolutional Neural Network.

Area of Science:

• Computer vision
• Medical imaging
• Artificial intelligence in medicine

Background:

• Video capsule endoscopy (VCE) is crucial for diagnosing gastrointestinal conditions.
• Deep learning offers potential to enhance VCE analysis, improving anomaly detection and reducing physician workload.
• Current computer-aided diagnosis (CADx) systems require high throughput and accuracy for clinical utility.

Purpose of the Study:

• To propose FocalConvNet, a novel deep learning model for VCE analysis.
• To classify small bowel anatomical landmarks and luminal findings efficiently.
• To enhance the accuracy and throughput of automated VCE diagnostic systems.

Main Methods:

• Developed FocalConvNet, integrating focal modulation with lightweight convolutional layers.
• Leveraged focal modulation for global context and global-local spatial interactions.
• Utilized convolutional blocks for hierarchical feature extraction and learning bias.

Main Results:

• FocalConvNet achieved a weighted F1-score of 0.6734, recall of 0.6373, and MCC of 0.2974 on the Kvasir-Capsule dataset.
• Outperformed state-of-the-art methods on a large-scale VCE dataset (44,228 frames, 13 classes).
• Achieved a high throughput of 148.02 images/second, demonstrating real-time clinical potential.

Conclusions:

• FocalConvNet shows significant promise for improving automated VCE analysis.
• The model's high throughput and accuracy support its potential for real-time clinical applications.
• This deep learning approach can aid physicians in faster and more accurate diagnosis of VCE findings.