Abstract
Local noise power spectra (NPS) have been commonly calculated to represent the noise properties of CT imaging systems, but their properties are significantly affected by the utilized calculation schemes. In this study, the effects of varied calculation parameters on the local NPS were analyzed, and practical suggestions were provided regarding the estimation of local NPS for clinical CT scanners. The uniformity module of a Catphan phantom was scanned with a Philips Brilliance 64 slice CT simulator with varied scanning protocols. Images were reconstructed using FBP and iDose4 iterative reconstruction with noise reduction levels 1, 3, and 6. Local NPS were calculated and compared for varied region of interest (ROI) locations and sizes, image background removal methods, and window functions. Additionally, with a predetermined NPS as a ground truth, local NPS calculation accuracy was compared for computer simulated ROIs, varying the aforementioned parameters in addition to ROI number. An analysis of the effects of these varied calculation parameters on the magnitude and shape of the NPS was conducted. The local NPS varied depending on calculation parameters, particularly at low spatial frequencies below ∼0.15 mm−1. For the simulation study, NPS calculation error decreased exponentially as ROI number increased. For the Catphan study the NPS magnitude varied as a function of ROI location, which was better observed when using smaller ROI sizes. The image subtraction method for background removal was the most effective at reducing low‐frequency background noise, and produced similar results no matter which ROI size or window function was used. The PCA background removal method with a Hann window function produced the closest match to image subtraction, with an average percent difference of 17.5%. Image noise should be analyzed locally by calculating the NPS for small ROI sizes. A minimum ROI size is recommended based on the chosen radial bin size and image pixel dimensions. As the ROI size decreases, the NPS becomes more dependent on the choice of background removal method and window function. The image subtraction method is most accurate, but other methods can achieve similar accuracy if certain window functions are applied. All dependencies should be analyzed and taken into account when considering the interpretation of the NPS for task‐based image quality assessment.PACS number(s): 87.57.C‐, 87.57.Q‐
Highlights
A minimum region of interest (ROI) size is recommended based on the chosen radial bin size and image pixel dimensions
As the ROI size decreases, the noise power spectrum (NPS) becomes more dependent on the choice of background removal method and window function
While there is little dependence of the NPS on background removal method for the 128-pixel ROI, as seen in Fig. 3(a), the calculated NPS is dependent on the background removal method for the smallest 32-pixel ROI, at low spatial frequencies
Summary
The noise power spectrum (NPS), as a more thorough noise descriptor than pixel standard deviation, describes both the magnitude and spatial frequency characteristics of image noise, which plays a critical role in analyzing and optimizing imaging system performance.[1,2,3,4,5] It is most often integrated with other metrics to assess image quality for specific tasks,(6-9) and has been commonly utilized in the development, characterization, optimization, and comparison of many new imaging technologies such as computed radiography,(10) digital mammogra phy,(11-12) storage phosphors for dental X-ray,(13) and other devices in a preclinical[14,15] and clinical[16,17] environment.There is, a fundamental limitation for the usage of the NPS for CT image noise assessment: the requirement of wide-sense stationary noise.[18]. Due to the intrinsic physics of the CT acquisition process, volumetric CT images violate this condition; this has been reported as a known problem.[18,19,20] NPS calculation methods should consider this limitation and attempt to acquire image data samples with more stationary noise properties by varying experimental conditions. The calculated NPS magnitude (i.e., noise variance) for the 32-pixel ROI increases as the ROIs are closer to the center of the image due to the attenuation properties of photons This verified the trend reported in a previous study,(20) and demonstrates the importance of utilizing smaller ROIs, which contain more information about the spatial variation of noise. The same trends are observed for both FBP and iDose reconstructions
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
