Marc Kachelriess PhD, PRESENTER: Nothing to Disclose

Abstract: HTML Purpose: Projection noise and image noise in CT mainly result from x-ray photon quantum noise. Often, quantum noise in CT is assumed to be Poisson distributed with a variance of N=N0 ·p where N0 is the number of incident quanta and p the probability for a photon to pass the object and to be detected. Our aim was to evaluate the validity of this simple approach and to propose and evaluate improved methods. Methods and Materials: We used (I) the physically correct polychromatic binomial noise model that sums the variances N0b ·pb ·(1-pb) over the energy bins b as the reference model. This reference was compared to (II) the simple Poisson model and (III) to a simple binomial noise model that is a pragmatic compromise between the two previous approaches and assumes N0·p·(1-p) to be the quantum noise variance. Simulations of projection and image noise using error propagation and reconstructions of a generalized water-equivalent oval CTDI phantom with object sizes ranging from 6 to 48 cm using tube voltages ranging from 30 to 200 kV were performed. The influence of data precorrection functions upon noise prediction was evaluated by in- and excluding a water precorrection into the error propagation. A simple polynomial precorrection function was used to linearize polychromatic water attenuation for a 120 kV spectrum. Quotients Q(x, y) of sigma images σ(x, y), giving image noise relative to the reference model, quantified the results. Results: The Poisson model is accurate only for high attenuation values (i.e., large object sizes). For objects below 10 cm maximum diameter the noise predicted by the Poisson approach is significantly higher than the true physical result. E.g., for a 6 cm oval the Poisson approach generates 30% more noise than the reference model. The proposed simple binomial model predicts noise for all simulated object sizes and tube spectra with errors below 1%. Taking into account the precorrection is inevitable: Otherwise significant deviations from the true image noise will be the case. We observed relative errors up to 14% even though our water precorrection was close to the identity transform. Conclusion: The simple Poisson model is valid only for large objects and high attenuation values. It should be replaced by the simple binomial model. The influence of precorrection functions must be taken into account. Only then, physically realistic results can be achieved. Especially for object diameters below 10 cm, as they occur in pediatric CT, in CT mammography or in micro CT, the improved accuracy is important when designing scan protocols for new applications.       Questions about this event email: marc.kachelriess@imp.uni-erlangen.de

Kachelriess PhD, M, Noise Prediction and Simulation in CT.  Radiological Society of North America 2003 Scientific Assembly and Annual Meeting, November 30 - December 5, 2003 ,Chicago IL. http://archive.rsna.org/2003/3100647.html