Noise2Noise-Based Self-Supervised Denoising Method for Single-Step Material Decomposition in Spectral CT

Photon-counting CT (PCCT) utilizes spectral data to generate material decomposition (MD) images, and single-step decomposition method offers distinct advantages by eliminating the need for image reconstruction, offering clinical benefits but facing noise amplification issues requiring denoising. The scarcity of clean reference images limits supervised (Noise2Clean) approaches. Noise2Noise circumvents this by training on paired noisy images with independent noise. Leveraging PCCT’s Poissondistributed photon counts across energy bins, this paper proposes binomial selection to split projection data into two low-dose scans with independent noise. Noise propagation analysis confirms inherited noise independence in decomposed images, validated through simulations and actual animal studies. Benchmarks show this Noise2Noise framework with binomial selection surpasses existing self-supervised techniques in noise suppression and feature preservation.