PFCM: Poisson Flow Consistency Models for Low-Dose CT Image Denoising.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Science

Background:

• X-ray computed tomography (CT) is crucial for medical diagnosis but involves ionizing radiation exposure.
• Optimizing CT image quality at reduced radiation doses is a significant research challenge.
• Deep generative models show promise for image denoising applications.

Purpose of the Study:

• Introduce Poisson Flow Consistency Models (PFCM), a novel deep generative model.
• Adapt PFCM for effective low-dose CT image denoising.
• Evaluate the performance and generalizability of PFCM in CT denoising.

Main Methods:

• Developed PFCM by generalizing consistency distillation to PFGM++.
• Utilized a tunable hyperparameter D to balance robustness and rigidity.
• Implemented a "task-specific" sampler that integrates low-dose CT images into the generative process.
• Assessed denoising performance using LPIPS, SSIM, and PSNR metrics.

Main Results:

• PFCM-based sampler achieved excellent denoising performance on the Mayo low-dose CT dataset.
• The model demonstrated robustness to noise mismatches inherent in low-dose CT.
• PFCM outperformed standard consistency models in low-dose CT denoising tasks.
• Effective denoising was shown on clinical images from a photon-counting system.

Conclusions:

• PFCM presents a robust and efficient framework for low-dose CT image denoising.
• The tunable hyperparameter D is critical for adapting the model to specific imaging challenges.
• PFCM demonstrates significant potential for improving medical imaging safety and quality.