Model-based iterative reconstruction (MBIR) methods based on maximum a posteriori (MAP) estimation have been
recently introduced to multi-slice CT scanners. The model-based approach has shown promising image quality
improvement with reduced radiation dose compared to conventional FBP methods, but the associated high computation
cost limits its widespread use in clinical environments. Among the various choices of numerical algorithms to optimize
the MAP cost function, simultaneous update methods such as the conjugate gradient (CG) method have a relatively high
level of parallelism to take full advantage of a new generation of many-core computing hardware. With proper
preconditioning techniques, fast convergence speeds of CG algorithms have been demonstrated in 3D emission and 2D
transmission reconstruction. However, 3D transmission reconstruction using preconditioned conjugate gradient (PCG)
has not been reported. Additional challenges in applying PCG in 3D CT reconstruction include the large size of clinical
CT data, shift-variant and incomplete sampling, and complex regularization schemes to meet the diagnostic standard of
image quality. In this paper, we present a ramp-filter based PCG algorithm for 3D CT MBIR. Convergence speeds of
algorithms with and without using the preconditioner are compared.
In a conventional X-ray CT system, where an object is scanned with a selected incident x-ray spectrum, or kVp, the
reconstructed images only approximate the linear X-ray attenuation coefficients of the imaged object at an effective
energy of the incident X-ray beam. The errors are primarily the result of beam hardening due to the polychromatic nature
of the X-ray spectrum. Modem clinical CT scanners can reduce this error by a process commonly referred to as spectral
calibration. Spectral calibration linearizes the measured projection value to the thickness of water. However, beam
hardening from bone and contrast agents can still induce shading and streaking artifacts and cause CT number
inaccuracies in the image.
In this paper, we present a dual kVp scanning method, where during the scan, the kVp is alternately switching between
target low and high preset values, typically 80kVp and 140 kVp, with a period less than 1ms. The measured projection
pairs are decomposed into the density integrals of two basis materials in projection space. The reconstructed density
images are further processed to obtain monochromatic attenuation coefficients of the object at any desired energy.
Energy levels yielding optimized monochromatic images are explored, and their analytical representations are derived.
Recently there has been significant interest in dual energy CT imaging with several acquisition methods being
actively pursued. Here we investigate fast kVp switching where the kVp alternates between low and high kVp
every view. Fast kVp switching enables fine temporal registration, helical and axial acquisitions, and full field
of view. It also presents several processing challenges. The rise and fall of the kVp, which occurs during the
view integration period, is not instantaneous and complicates the measurement of the effective spectrum for low
and high kVp views. Further, if the detector digital acquisition system (DAS) and generator clocks are not fully
synchronous, jitter is introduced in the kVp waveform relative to the view period.
In this paper we develop a method for estimation of the resulting spectrum for low and high kVp views. The
method utilizes static kVp acquisitions of air with a small bowtie filter as a basis set. A fast kVp acquisition of
air with a small bowtie filter is performed and the effective kVp is estimated as a linear combination of the basis
vectors. The effectiveness of this method is demonstrated through the reconstruction of a water phantom acquired
with a fast kVp acquisition. The impact of jitter due to the generator and detector DAS clocks is explored via
simulation. The error is measured relative to spectrum variation and material decomposition accuracy.
Dual energy CT cardiac imaging is challenging due to cardiac motion and the resolution requirements of clinical
applications. In this paper we investigate dual energy CT imaging via fast kVp switching acquisitions of a novel
dynamic cardiac phantom. The described cardiac phantom is realistic in appearance with pneumatic motion control
driven by an ECG waveform.
In the reported experiments the phantom is driven off a 60 beats per minute simulated ECG waveform. The cardiac
phantom is inserted into a phantom torso cavity. A fast kVp switching axial step and shoot acquisition is detailed. The
axial scan time at each table position exceeds one heart cycle so as to enable retrospective gating. Gating is performed
as a mechanism to mitigate the resolution impact of heart motion.
Processing of fast kVp data is overviewed and the resulting kVp, material decomposed density, and monochromatic
reconstructions are presented. Imaging results are described in the context of potential clinical cardiac applications.
Linear discriminate analysis (LDA) is applied to dual kVp CT and used for tissue characterization. The
potential to quantitatively model both malignant and benign, hypo-intense liver lesions is evaluated by analysis of
portal-phase, intravenous CT scan data obtained on human patients. Masses with an a priori classification are mapped
to a distribution of points in basis material space. The degree of localization of tissue types in the material basis space
is related to both quantum noise and real compositional differences. The density maps are analyzed with LDA and
studied with system simulations to differentiate these factors. The discriminant analysis is formulated so as to
incorporate the known statistical properties of the data. Effective kVp separation and mAs relates to precision of tissue
localization. Bias in the material position is related to the degree of X-ray scatter and partial-volume effect.
Experimental data and simulations demonstrate that for single energy (HU) imaging or image-based decomposition
pixel values of water-like tissues depend on proximity to other iodine-filled bodies. Beam-hardening errors cause a
shift in image value on the scale of that difference sought between in cancerous and cystic lessons. In contrast,
projection-based decomposition or its equivalent when implemented on a carefully calibrated system can provide
accurate data. On such a system, LDA may provide novel quantitative capabilities for tissue characterization in dual
In 3rd generation CT systems projection data, generated by X-rays emitted from a single source and passing
through the imaged object, are acquired by a single detector covering the entire field of view (FOV). Novel
CT system architectures employing distributed sources [1,2] could extend the axial coverage, while
removing cone-beam artifacts and improving spatial resolution and dose. The sources can be distributed in
plane and/or in the longitudinal direction. We investigate statistical iterative reconstruction of multi-axial
data, acquired with simulated CT systems with multiple sources distributed along the in-plane and
longitudinal directions. The current study explores the feasibility of 3D iterative Full and Half Scan
reconstruction methods for CT systems with two different architectures. In the first architecture the sources
are distributed in the longitudinal direction, and in the second architecture the sources are distributed both
longitudinally and trans-axially. We used Penalized Weighted Least Squares Transmission Reconstruction
(PWLSTR) and incorporated a projector-backprojector model matching the simulated architectures. The
proposed approaches minimize artifacts related to the proposed geometries. The reconstructed images show
that the investigated architectures can achieve good image quality for very large coverage without severe
Three-dimensional iterative reconstruction of large CT data sets poses several challenges in terms of the associated
computational and memory requirements. In this paper, we present results obtained by implementing
a computational framework for reconstructing axial cone-beam CT data using a cluster of inexpensive dualprocessor
PCs. In particular, we discuss our parallelization approach, which uses POSIX threads and message
passing (MPI) for local and remote load distribution, as well as the interaction of that load distribution with
the implementation of ordered subset based algorithms. We also consider a heuristic data-driven 3D focus of
attention algorithm that reduces the amount of data that must be considered for many data sets. Furthermore,
we present a modification to the SIRT algorithm that reduces the amount of data that must be communicated
between processes. Finally, we introduce a method of separating the work in such a way that some computation
can be overlapped with the MPI communication thus further reducing the overall run-time. We summarize the
performance results using reconstructions of experimental data.