HQG-Net: Unpaired Medical Image Enhancement With High-Quality Guidance.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Unpaired medical image enhancement (UMIE) aims to improve low-quality (LQ) medical images without paired training data.
• Existing UMIE methods, often based on GANs, struggle to explicitly leverage high-quality (HQ) information, leading to artifacts and distortions.

Purpose of the Study:

• To propose a novel UMIE approach that directly encodes HQ cues into the enhancement process.
• To improve the visual quality and structural integrity of enhanced medical images.
• To ensure enhanced images are favorable for downstream diagnostic tasks.

Main Methods:

• A variational approach to model the joint distribution between LQ and HQ image domains.
• Explicitly inserting HQ image features into the enhancement network via a variational normalization module.
• Adversarial training with a discriminator, content-aware loss (wavelet-based pixel and multi-encoder feature levels), and a bi-level learning scheme for cooperative optimization with downstream tasks.

Main Results:

• The proposed method effectively guides LQ medical image enhancement using HQ cues.
• The generated HQ images exhibit improved visual appeal and reduced structural distortions.
• The enhanced images demonstrate superior performance in downstream tasks compared to existing methods.
• Experimental validation on three medical datasets confirmed the method's effectiveness.

Conclusions:

• The novel UMIE approach successfully integrates HQ information to enhance LQ medical images.
• The method provides visually appealing and functionally useful enhanced images for clinical applications.
• The bi-level learning scheme optimizes both enhancement and downstream task performance, showcasing the practical utility of the approach.