Author: Xinhui Duan, Roderick W. McColl, Mi-Ae Park, Liqiang Ren, Gary Xu, Kuan Zhang, Yue Zhang 👨🔬
Affiliation: UT Southwestern Medical Center, Department of Radiology, UT Southwestern Medical Center, Imaging Services, UT Southwestern Medical Center 🌍
Purpose:
Image-based deep-learning noise-reduction techniques have been developed for photon-counting CT (PCCT) to improve image quality with reduced radiation dose. The denoising strength is typically governed by a default noise weighting factor (α) during training, representing the added noise level between noisy inputs and reference. Although αa is empirically set (e.g., 2 for a coronary CT angiography study), the optimal denoising strength can vary by patient size and diagnostic task, which can be determined using input and target. In this work, we developed an automated method to detect the noise level between noisy and less-noise images.
Methods:
Twenty-eight abdomen exams were acquired using PCCT (NAEOTOM Alpha, Siemens Healthineers). Virtual monoenergetic images (VMIs) at 50 keV were reconstructed with both filtered back-projection and iterative reconstruction at strength level 3 (QIR-3). Subtracting these reconstructions yielded a noise map, which was spatially decoupled and normalized to improve generalizability. This noise map, weighted by factor α ranging from 0.8 to 5.0 in 0.1 increments, was inserted into the IR images with 5 mm slice thickness to produce noisy images. Pairs of noisy and IR images were concatenated as input data, and α served as the label, resulting in a dataset of 1204 cases. A regression model based on 3D convolutional residual network was designed to map PCCT image pairs to α. Performance was evaluated using the mean square error (MSE).
Results:
The noise-level detector accurately identified the relative noise in PCCT images, achieving MSE values of 0.11 on the validation set (N=241) and 0.20 on the testing set (N=241) between the model-predicted α* and the true α. In comparison, predicting a constant α^ yielded an MSE of 1.54.
Conclusion:
These findings demonstrate the feasibility of automatically estimating noise levels in PCCT images, a significant step towards personalized noise- and dose-reduction strategies in clinical practice.