Generalizable medical image enhancement using structure-preserved diffusion models.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Clinical medical image quality is crucial for accurate diagnosis by clinicians and AI.
• Existing Generative Adversarial Network (GAN)-based enhancement methods face challenges like artifacts and training instability.
• Diffusion models offer superior image generation but struggle with medical data collection and domain gaps, particularly in preserving fine structures.

Purpose of the Study:

• To develop a generalizable medical image enhancement method that preserves fine structures.
• To address limitations of current diffusion models in medical image enhancement.
• To improve the diagnostic utility of low-quality clinical images.

Main Methods:

• Proposed a generalizable medical image enhancement using structure-preserved diffusion models (GEDM).
• Utilized joint supervision from enhancement and segmentation tasks for improved structure preservation and generalizability.
• Employed synthetic data for paired training data collection and the Laplace transform to reduce domain gaps and incorporate multi-scale information.

Main Results:

• GEDM demonstrated superior performance in image enhancement compared to state-of-the-art methods.
• The method effectively preserved fine structures in enhanced medical images.
• Enhanced image quality led to improved performance in subsequent medical analysis tasks.

Conclusions:

• GEDM offers a robust solution for medical image enhancement, outperforming existing techniques.
• The joint enhancement and segmentation approach effectively preserves structural details.
• The proposed method shows significant potential for improving clinical diagnostics and AI-assisted medical analysis.