A Dose-Leveled Parametric NPS Modeling Library for Photon-Counting CT
Abstract
Purpose
Photon-counting CT (PCCT) denoising seeks to approximate high-dose image quality from lower-dose acquisitions. A key challenge for patient-specific denoising with precise control of denoising strength is the lack of an available ground-truth “target” image whose noise characteristics match a desired dose condition. In this work, we developed a dose-leveled Noise Power Spectrum (NPS) modeling library that predicts CT noise magnitude and frequency content continuously across dose levels, reconstruction strengths, and image types without requiring repeated acquisitions at each condition.
Methods
An ACR CT accreditation phantom was scanned on a PCCT system (NAEOTOM Alpha, Siemens) using a routine abdomen-pelvis protocol. CTDIvol were varied from 1-12 mGy. Reconstructions included low-energy threshold images (T3D) and virtual monoenergetic images (VMIs) at 50 and 70keV, each reconstructed using quantum iterative reconstruction (QIR-3) and filtered back-projection–like reconstruction (QIR-0) with a body kernel (Br44). Noise was measured in the uniformity module using multiple square regions of interest to compute the 2D NPS. We proposed an analytical parametric form to mimic the NPS profile. Model parameters were fit separately for each reconstruction condition using measurements from odd CTDIvol levels, and performance was evaluated on held-out even CTDIvol levels.
Results
The proposed predictor accurately captured both dose-dependent scaling and NPS shape across all reconstruction conditions. When trained on odd dose levels, the model generalized well to independent even dose levels. Predicted NPS closely matched measured NPS, yielding low mean squared error (10-3-10-2) and high coefficients of determination (R2>0.99). No meaningful domain shift was observed between training and testing dose levels.
Conclusion
This parameterized NPS modeling framework enables robust, continuous prediction of CT noise characteristics across dose and reconstruction settings. The resulting library provides a practical tool for dose-aware image quality assessment, protocol harmonization, and simulation-based PCCT studies, reducing the need for repeated acquisitions across multiple dose levels.