WE-D-18A-06: Estimating Local Noise Power Spectrum From a Few CT Scans




The traditional way to estimate CT NPS is by averaging the power spectrum of many noisy scans. When only a few scans are available, regions of interests are often extracted from different locations to obtain sufficient samples to estimate NPS, thus ignoring the nonstationarity of CT noise. This approach to estimating NPS does not accurately reflect the location-variant characteristics of CT noise. The purpose of this work is to develop a method to estimate local NPS from only a few CT scans.


As a result of the FBP reconstruction algorithm, the CT NPS presents the following radial property: the shape of the radial profile of the NPS is almost constant, determined by the reconstruction filter, and the magnitude varies with angle depending on the object attenuation map. Therefore, a shape function and an angular magnitude function are sufficient to describe the CT NPS. Based on this property, the dimensionality of the NPS is greatly reduced and we are able to derive a radial NPS method to estimate the NPS from only a few scans.


We applied the radial NPS method to simulated CT scans of a nonuniform object. The results showed that the local NPS estimated from only 6 scans using the radial method was very close to the NPS estimated using the traditional method from 400 scans, according to normalized mean squared error (NMSE) and the signal detectability based on an NPS-prewhitening Hoteling model observer. We also applied this method to physical phantom scans. Good accuracy was again achieved from only 6 scans using the radial NPS method.


The radial NPS method was shown to be accurate and efficient in estimating the local NPS of FBP reconstructed CT images. This method is readily extendable to estimating the NPS of helical CT scans.