Iterative material decomposition for spectral CT using self-supervised Noise2Noise prior.

Area of Science:

• Medical Imaging
• Image Processing
• Computational Science

Background:

• Spectral computed tomography (CT) offers material decomposition capabilities beyond conventional CT.
• Image noise is a significant challenge in spectral CT material decomposition, particularly under dose limitations.
• Existing denoising methods may require external training data or struggle with noise magnification.

Purpose of the Study:

• To develop an iterative, one-step inversion material decomposition algorithm for spectral CT.
• To integrate a self-supervised Noise2Noise prior for effective denoising without external training data.
• To enhance image quality and diagnostic utility of spectral CT material decomposition.

Main Methods:

• An iterative one-step inversion algorithm was designed, directly estimating material images from projection data.
• A Noise2Noise prior, based on self-supervised learning, was incorporated for image denoising.
• The algorithm utilized separable quadratic surrogate (SQS) for data consistency and Adam for network optimization.

Main Results:

• The proposed method demonstrated significant noise suppression in simulated, preclinical, and clinical spectral CT data.
• Quantitative analysis confirmed superior structure detail recovery compared to other methods.
• The self-supervised approach eliminated the need for external training datasets.

Conclusions:

• The developed iterative algorithm with a Noise2Noise prior effectively addresses noise in spectral CT material decomposition.
• This self-supervised method offers a promising approach for improving diagnostic accuracy in spectral CT imaging.
• The technique shows potential for application across various spectral CT platforms, including photon-counting and dual-energy CT.