Alexander Zamyatin PhD, Presenter: Employee, Toshiba Corporation

Satoru Nakanishi MS, Abstract Co-Author: Employee, Toshiba Corporation

Yi Fan PhD, Abstract Co-Author: Employee, Toshiba Corporation

Kurt Schultz RT, Abstract Co-Author: Employee, Toshiba Corporation

Validation of low-dose techniques such as iterative reconstruction requires an accurate model of CT data. We propose and evaluate a noise model based on energy-dependent compound Poisson statistics and evaluate its accuracy with low dose CT data.

Our noise model is based on measuring mean and variance tables in unprocessed CT data by repeating scans of air and various water cylinder sizes. Measured noise variance is offset by electronic noise variance to obtain the intensity signal variance. From these measurements we estimate a variance gain, or variance to mean ratio, which, according to Poisson noise model, must be constant for all measurements. We found that measured noise variance gain is linearly proportional to the effective energy of the x-ray beam, as expected from the compound Poisson model. Effective energy varies from channel to channel due to beam hardening in bowtie and object. Due to high variation of effective energy within one view, the Poisson model is less accurate than the compound Poisson model for estimation of the noise variance. In the patient scans the effective energy at each detector bin is not known, but can be estimated based on measured intensity and tube voltage. Knowing the system bowtie shape is not necessary; if necessary, it can be estimated from the noise variance measurements. For evaluation we used CT phantoms (Catphan, TOS), and anthropomorphic phantoms at various kV and mA settings.

Noise model accuracy is measured by comparing image noise in real and simulated low dose CT scans. Simulated low dose scans are obtained from regular dose CT scans by adding noise according to our model. For cylindrical phantoms, both Poisson and compound Poisson models have good accuracy. Poisson model is accurate when calibration scan uses a water cylinder of same size as the phantom size; however, for anthropomorphic phantoms Poisson noise model results in larger differences with real noise, depending on phantom size and position. Since compound Poisson model does not assume cylindrical object shape, its accuracy for anthropomorphic phantoms is as good as with cylindrical phantoms.

For accurate simulation of low dose clinical CT data we suggest using Compound Poisson model. 

Accurate CT noise model is necessary for reducing patient dose in clinical CT. Its applications are model-based iterative reconstruction and low dose algorithm validation.

Zamyatin, A, Nakanishi, S, Fan, Y, Schultz, K, Energy-dependent Noise Modeling and Application to Low-Dose CT Imaging.  Radiological Society of North America 2012 Scientific Assembly and Annual Meeting, November 25 - November 30, 2012 ,Chicago IL. http://archive.rsna.org/2012/12043864.html