BEST IN PHYSICS (IMAGING): A Patient-Specific, Target NPS-Driven Noise Insertion Framework for Photon-Counting CT Denoising
Abstract
Purpose
Photon-counting CT (PCCT) denoising requires paired images with different noise levels for supervised training, but acquiring such data is rarely feasible due to increased radiation exposure and registration challenges. As a result, low-dose–like data are generated by inserting noise into high-dose images, although existing methods rely on empirical weighting factors that can misestimate noise and fail to reproduce the correct NPS. To address this, we developed a patient-specific noise insertion framework that enables precise control of the noise power spectrum, producing realistic low-dose–like images for PCCT denoising.
Methods
An ACR CT phantom and routine abdomen–pelvis patient scans were acquired on a clinical PCCT (NAEOTOM Alpha, Siemens). Images at CTDIvol of 8 mGy and 4 mGy were reconstructed using quantum iterative reconstruction (QIR3) and filtered back-projection QIR0 with a Br44 kernel. Residual noise obtained from QIR3–QIR0 subtraction was spatially decoupled using phase randomization. A target-NPS was defined from the difference between low-dose QIR0 and high-dose QIR3 images, and the decoupled noise was spectrally reshaped and added to high-dose QIR3 images to generate low-dose IR0–like images. Performance was evaluated by comparing NPS with corresponding real low-dose QIR0 images.
Results
Phase randomization effectively removed structural components while maintaining NPS characteristics. Synthesized images showed high spectral agreement with real low-dose QIR0 images, with a low MSE of 0.02 between NPS curves. Dominant noise frequency content was preserved, with peak spatial frequencies of 2.51 cm-1 for real and 2.36 cm-1 for synthesized images, corresponding to a peak shift of −0.15 cm-1. Synthesized images remained distinguishable from lose-dose QIR0 images, confirming accurate noise insertion.
Conclusion
The proposed patient-specific, target NPS–driven noise insertion framework enables controlled synthesis of low-dose-like PCCT images using high-dose data alone. This approach provides a practical tool for CT noise modeling, image-quality assessment, and data augmentation when multi-dose acquisitions are limited.