Matthias Baer DiplPhys, Presenter: Nothing to Disclose

Marc Kachelriess PhD, Abstract Co-Author: Nothing to Disclose

To provide a fast and efficient scatter correction for clinical and flat detector CT that can, potentially, be used to predict scatter distributions in real-time.

CT projection data are often impaired by scatter. While scatter is suppressed to some extent by anti scatter grids, it still reduces the accuracy of quantitative evaluations and produces significant artifacts in flat detector CT, due to rather large cone angles and frequently missing anti scatter grids. Monte Carlo-based scatter correction is considered to be very accurate for scatter prediction and removal. However accurate scatter prediction requires the calculation of many photon histories resulting in a computational complexity too high for routine use. Convolution-based scatter correction methods have far less computational needs since scatter is estimated by scaling and low pass filtering measured intensities. Their drawbacks are open parameters which must be calibrated in advance based on reference objects and thus they are not patient-specific. We propose to combine a real-time patient-specific Monte Carlo scatter prediction, based on a few photon histories, with a convolution approach. The Monte Carlo output is used to calibrate the parameters of the convolution model on a patient- and projection-specific basis. Our hybrid approach can be thought as regularizing noisy Monte Carlo estimates with a convolution-based model. The method is validated using simulated data and data of a flat detector CT system (True Beam, Varian, Palo Alto, USA).  

Regularizing with a physical convolution model allows reducing the number of photon histories by a factor of 50 to 100 without loss in scatter prediction accuracy. Thereby it was sufficient to calibrate the convolution model at an angular increment of 22.5°. Compared to a low variance Monte Carlo scatter correction CT value deviations of 105 HU / 38 HU (measurements / simulations) in uncorrected images were reduced to 1 HU with HSC. The total runtime of HSC was 50 s to predict the scatter intensity of a 5123 cone-beam CT scan.

HSC allows to predict scatter almost as accurately as a low variance Monte Carlo at roughly 1% to 2% of the Monte Carlo computation time.

Especially in flat panel CT scatter significantly impairs image quality. Our method is a new option to strongly reduce scatter artifacts in real time and has the potential to be used in daily routine.

Baer, M, Kachelriess, M, Hybrid Scatter Correction (HSC) for Diagnostic CT and for Flat Detector CT  .  Radiological Society of North America 2012 Scientific Assembly and Annual Meeting, November 25 - November 30, 2012 ,Chicago IL. http://archive.rsna.org/2012/12024655.html