Noise suppression in photon-counting computed tomography using unsupervised Poisson flow generative models.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Science

Background:

• Deep learning (DL) excels at computed tomography (CT) image denoising but often requires supervised training with hard-to-obtain paired data.
• Diffusion models offer unsupervised solutions for inverse problems through posterior sampling, but iterative solvers lead to high numbers of function evaluations (NFE).

Purpose of the Study:

• To develop a novel, unsupervised deep learning technique for photon-counting CT image denoising.
• To enable single-step (NFE=1) posterior sampling for CT denoising using generative models.

Main Methods:

• Extended unsupervised inverse problem solving to Poisson flow generative models (PFGM)++.
• Implemented a single-step sampler by hijacking and regularizing the PFGM++ sampling process.
• Incorporated posterior sampling using diffusion models as a specific case within the PFGM++ framework.

Main Results:

• Achieved single-step (NFE=1) image denoising for photon-counting CT.
• Demonstrated significant performance gains due to the robustness of the PFGM++ framework.
• Showcased competitive results against supervised, unsupervised, and non-DL denoising methods on clinical data.

Conclusions:

• The proposed PFGM++ based method offers an effective unsupervised approach for single-step CT image denoising.
• This technique provides a robust and efficient alternative to traditional supervised and iterative unsupervised methods.
• The approach shows promise for improving image quality in photon-counting CT applications.