Improving Statistical Image Reconstruction for Cardiac X-ray Computed Tomography.

Abstract

Technological advances in CT imaging pose new challenges such as increased X-ray radiation dose and complexity of image reconstruction. Statistical image reconstruction methods use realistic models that incorporate the physics of the measurements and the statistical properties of the measurement noise, and they have potential to provide better image quality and dose reduction compared to the conventional filtered back-projection (FBP) method. However, statistical methods face several challenges that should be addressed before they can replace the FBP method universally. In this thesis, we develop various methods to overcome these challenges of statistical image reconstruction methods.

Rigorous regularization design methods in Fourier domain were proposed to achieve more isotropic and uniform spatial resolution or noise properties. The design framework is general so that users can control the spatial resolution and the noise characteristics of the estimator. In addition, a regularization design method based on the hypothetical geometry concept was introduced to improve resolution or noise uniformity. Proposed designs using the new concept effectively improved the spatial resolution or noise uniformity in the reconstructed image. The hypothetical geometry idea is general enough to be applied to other scan geometries.

Statistical weighting modification, based on how much each detector element affects insufficiently sampled region, was proposed to reduce the artifacts without degrading the temporal resolution within the region-of-interest (ROI). Another approach using an additional regularization term, that exploits information from the prior image, was investigated. Both methods effectively removed short-scan artifacts in the reconstructed image.

We accelerated the family of ordered-subsets algorithms by introducing a double surrogate so that faster convergence speed can be achieved. Furthermore, we present a variable splitting based algorithm for motion-compensated image reconstruction (MCIR) problem that provides faster convergence compared to the conjugate gradient (CG) method. A sinogram-based motion estimation method that does not require any additional measurements other than the short-scan amount of data was introduced to provide decent initial estimates for the joint estimation.

Proposed methods were evaluated using simulation and real patient data, and showed promising results for solving each challenge. Some of these methods can be combined to generate more complete solutions for CT imaging.