Clinically Generalizable Low-Dose CT Denoising for Pediatric Imaging via Enhanced Diffusion Posterior Sampling.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Reducing radiation dose in CT scans is crucial for patient safety, especially in pediatric PET/CT imaging.
• Deep learning methods for low-dose CT (LDCT) to normal-dose CT (NDCT) denoising often require paired data, which is clinically challenging to obtain.
• Existing methods struggle with generalizability to real data or preserving structural integrity.

Purpose of the Study:

• To develop a novel deep learning framework for robust LDCT denoising in PET/CT imaging.
• To enhance the generalizability and structural fidelity of denoised CT images.
• To reduce radiation exposure in total-body PET/CT scans without compromising diagnostic quality.

Main Methods:

• Proposed an enhanced diffusion posterior sampling (E-DPS) framework combining a U-Net denoiser with an unconditional diffusion model.
• The U-Net provides structural constraints from simulated LDCT-NDCT pairs, while the diffusion model learns the NDCT prior distribution.
• Implemented an intermediate-stage initialization strategy to reduce sampling steps during inference.

Main Results:

• E-DPS achieved superior performance on simulated LDCT datasets, demonstrating significant PSNR gains (+5.2% and +4.3%) at unseen dose levels.
• The method exhibited strong zero-shot generalizability on real LDCT images, effectively suppressing noise while preserving anatomical details.
• Outperformed state-of-the-art approaches in both simulated and real-world denoising scenarios.

Conclusions:

• The E-DPS framework offers a robust and clinically viable solution for LDCT denoising in total-body PET/CT.
• This approach effectively balances noise reduction with the preservation of crucial anatomical information.
• Highlights the potential for reduced radiation exposure in medical imaging through advanced AI techniques.