EndoL2H

EndoL2H enhances the resolution of wireless capsule endoscopy images to improve visualization and support diagnosis and characterization of small bowel pathologies.

Key Features:

• Super-Resolution Framework: Learns a mapping from low-resolution capsule endoscopic images to high-resolution counterparts to overcome limited camera resolution.
• Advanced Neural Network Architecture: Integrates conditional adversarial networks with a spatial attention block and achieves upscaling factors of 8×, 10×, and 12×.
• Quantitative Validation: Evaluated against Deep Back-Projection Networks (DBPN), Deep Residual Channel Attention Networks (RCAN), and Super Resolution Generative Adversarial Network (SRGAN) with superior quantitative and qualitative performance reported.
• Clinical Relevance: Image quality improvements were assessed by Mean Opinion Score (MOS) tests with 30 gastroenterologists for potential diagnostic impact.

Scientific Applications:

• Cross-system applicability: Applicable to images from various endoscopic capsule systems for improved image resolution.
• Polyp detection and characterization: Enhances visualization of polyps and other small bowel pathologies to aid detection and characterization.
• Support for computational diagnostics: Produces higher-resolution inputs to support automated and computational approaches in gastroenterology.

Methodology:

Applies deep learning to transform low-resolution endoscopic images into high-resolution outputs using conditional adversarial networks combined with spatial attention mechanisms.