Low-dose CT imaging using a regularization-enhanced efficient diffusion probabilistic model

Area of Science:

• Medical Imaging
• Computational Radiology
• Artificial Intelligence in Medicine

Background:

• Low-dose CT (LDCT) reduces radiation exposure but increases image noise, degrading quality for clinical tasks.
• Current denoising methods are often too slow for real-time clinical use due to extensive diffusion steps.

Purpose of the Study:

• Propose a regularization-enhanced efficient diffusion probabilistic model (RE-EDPM) for rapid, high-fidelity LDCT denoising.
• Incorporate residual guidance and hybrid regularization to maintain anatomical and diagnostic quality.

Main Methods:

• RE-EDPM uses a residual-shifting mechanism and a Swin-based U-Net with four reverse diffusion steps.
• A composite loss function (pixel, perceptual, TV regularization) suppresses noise while preserving fine details.
• Evaluated on public LDCT benchmarks using SSIM, PSNR, and VIFp, compared against state-of-the-art methods.

Main Results:

• Achieved high SSIM, PSNR, and VIFp scores on chest and abdominal LDCT datasets.
• Demonstrated superior noise suppression and structural fidelity compared to other methods.
• Processed two 512x512 slices in ~0.25s, enabling near-real-time application.

Conclusions:

• RE-EDPM provides effective LDCT denoising with minimal inference time.
• Balances noise reduction and anatomical preservation for clinical deployment.
• Suitable for real-time medical imaging and enhancing low-quality scans.