A Dual-Generalization Low-Light Enhancement Framework for Capsule Endoscopy Image Restoration and Segmentation

Area of Science:

• Medical Imaging
• Deep Learning
• Gastroenterology

Background:

• Deep learning aids gastrointestinal (GI) disease diagnosis via wireless capsule endoscopy (WCE).
• WCE images often suffer low-light degradation and varying details, impacting diagnostic accuracy.
• Existing methods struggle with diverse brightness and detail levels in WCE images.

Purpose of the Study:

• To develop a dual-generalization framework for enhancing low-light WCE images.
• To improve both image restoration and segmentation accuracy for WCE analysis.
• To enable robust diagnosis across varying image brightness and detail levels.

Main Methods:

• Proposed a dual-generalization framework incorporating Image Guidance and Laplacian Fusion Module (IGLFM), Brightness Level Generalization Module (BLGM), and Wavelet Segmentation Generalization Module (WSGM).
• IGLFM and BLGM focus on restoring low-light images across different brightness levels.
• WSGM enhances segmentation accuracy by generalizing varying image detail levels.

Main Results:

• The framework achieved significant performance gains in image restoration (e.g., 4.70 dB PSNR on Kvasir-Capsule).
• Demonstrated improved segmentation accuracy, with WSGM boosting mIoU by up to 4.7% on the RLE dataset.
• Outperformed state-of-the-art methods in handling varying brightness and detail levels.

Conclusions:

• The proposed dual-generalization framework effectively addresses low-light degradation and detail variations in WCE images.
• The method enhances both image quality and segmentation accuracy, crucial for accurate GI disease diagnosis.
• This approach offers a significant advancement for automated analysis of WCE imagery.