Dynamic contrast-enhanced (DCE) MRI is an important tool for the detection and characterization of primary and recurring prostate cancer. Advanced reconstruction strategies (e.g., sparse or low-rank regression) provide improved depiction of contrast dynamics and pharmacokinetic parameters; however, the high computation cost of reconstructing 4D (3D+time, 50+ frames) datasets typically inhibits their routine clinical use. Here, a novel alternating direction method-of-multipliers (ADMM) optimization strategy is described that enables these methods to be executed in ∠5 minutes, and thus within the standard clinical workflow. After overviewing the mechanics of this approach, high-performance implementation strategies will be discussed and demonstrated through clinical cases.
In many clinical MRI applications, not one but a series of images is acquired. Techniques that promote intra- and inter-image sparsity have recently emerged as powerful strategies for accelerating MRI applications; however, sparsity alone cannot always describe the complex relationships that exist between images in these series. In this paper, we will discuss the modeling of higher-dimensional MRI signals as matrices and tensors, and why promoting these signals to be low-rank (and, specifically, locally low-rank) can effectively identify and exploit these complex relationships. Example applications including training-free dynamic and calibrationless parallel MRI will be demonstrated.
Radiation dose from CT scans is an increasing health concern in the practice of radiology. Higher dose scans can
produce clearer images with high diagnostic quality, but may increase the potential risk of radiation-induced cancer or
other side effects. Lowering radiation dose alone generally produces a noisier image and may degrade diagnostic
performance. Recently, CT dose reduction based on non-local means (NLM) filtering for noise reduction has yielded
promising results. However, traditional NLM denoising operates under the assumption that image noise is spatially
uniform noise, while in CT images the noise level varies significantly within and across slices. Therefore, applying NLM
filtering to CT data using a global filtering strength cannot achieve optimal denoising performance. In this work, we
have developed a technique for efficiently estimating the local noise level for CT images, and have modified the NLM
algorithm to adapt to local variations in noise level. The local noise level estimation technique matches the true noise
distribution determined from multiple repetitive scans of a phantom object very well. The modified NLM algorithm
provides more effective denoising of CT data throughout a volume, and may allow significant lowering of radiation
dose. Both the noise map calculation and the adaptive NLM filtering can be performed in times that allow integration
with the clinical workflow.
We have recently developed a locally-adaptive method for noise control in CT based upon bilateral filtering.
Different from the previous adaptive filters, which were locally adaptive by adjusting the filter strength according to
local photon statistics, our use of bilateral filtering in projection data incorporates a practical CT noise model and takes
into account the local structural characteristics, and thus can preserve edge information in the projection data and
maintain the spatial resolution. Despite the incorporation of the CT noise model and local structural characteristics in the
bilateral filtering, the noise-resolution properties of the filtered image are still highly dependent on predefined
parameters that control the weighting factors in the bilateral filtering. An inappropriate selection of these parameters may
result in a loss of spatial resolution or an insufficient reduction of noise. In this work, we employed an adaptive strategy
to modulate the bilateral filtering strength according to the noise-equivalent photon numbers determined from each
projection measurement. We applied the proposed technique to head/neck angiographic CT exams, which had highly non-uniform attenuation levels during the scan. The results demonstrated that the technique can effectively reduce the noise and streaking artifacts caused by high attenuation, while maintaining the reconstruction accuracy in less attenuating regions.
In medical imaging, image background is often defined by zero signal. Moreover, in medical images the background
area - or conversely, the spatial support (the extent of the non-zero part of the image) - is often known a priori or can be
easily estimated. For example, support information can be estimated from the low-resolution "scout" images typically
acquired during pre-scan localization in both MRI and CT. In dynamic scans, object support in a single time-frame is
often obtainable from a prior time frame, or from a composite image formed from data from multiple time frames. In this work, incorporation of either complete or partial a priori knowledge of object spatial support into the compressive
sensing (CS) framework is investigated. Following development of the augmented reconstruction model, examples of
support-constrained CS reconstruction of phantom and MR images under both exact and inexact support definitions are
given. For each experiment, the straightforward incorporation of the proposed spatial support constraint into the standard CS model was shown to both significantly accelerate reconstruction convergence and yield a lower terminal RMSE compared to a conventional CS reconstruction. The proposed augmented reconstruction model was also shown to be robust to inaccuracies in the estimated object support.
Proc. SPIE. 7622, Medical Imaging 2010: Physics of Medical Imaging
KEYWORDS: Compressed sensing, Computed tomography, Image quality, Reconstruction algorithms, In vivo imaging, Data acquisition, Image restoration, Computer simulations, Signal attenuation, Signal to noise ratio
The purpose of this paper is
to present a new image reconstruction algorithm for dynamic data, termed non-convex prior
image constrained compressed sensing (NC-PICCS). It generalizes the prior image constrained compressed sensing
(PICCS) algorithm with the use of non-convex priors. Here, we concentrate on perfusion studies using computed
tomography examples in simulated phantoms (with and without added noise) and in vivo data, to show how the NC-PICCS
method holds potential for dramatic reductions in radiation dose for time-resolved CT imaging. We show that NC-PICCS can provide additional undersampling compared to conventional convex compressed sensing and PICCS, as well as, faster convergence under a quasi-Newton numerical solver.
Optimal noise control is critical for dose reduction in CT. In this work, we investigated the use of a locally-adaptive
method for noise reduction in low-dose CT. This method is based upon bilateral filtering, which smoothes the projection
data using a weighted average in a local neighborhood, where the weights are determined according to both the spatial
proximity and intensity similarity between the center pixel and the neighboring pixels. This filtering is locally adaptive
and can preserve important edge information in the sinogram, thus without significantly sacrificing the spatial resolution.
It is closely related to anisotropic diffusion, but is significantly faster. More importantly, a CT noise model can be
readily incorporated in the filtering and denoising process. We have evaluated the noise-resolution properties of the
bilateral filtering in a phantom study and a preliminary patient study with contrast-enhanced abdominal CT exams. The
results demonstrated that bilateral filtering can achieve a better noise-resolution tradeoff than a series of commercial
reconstruction kernels. This improvement on noise-resolution properties can be used for improving the image quality in
low-dose CT and can also be translated to substantial dose reduction.
A novel method for highly-undersampled Magnetic Resonance Image (MRI) reconstruction is presented. One
of the principal challenges faced in clinical MR imaging is the fundamental linear relation between net exam
duration and admissible spatial resolution. Increased scan duration diminishes patient comfort while increasing
the risk of susceptibility to motion artifact and limits the ability to depict many physiological events at high
temporal rates. With the recent development of Compressive Sampling theory, several authors have successfully
demonstrated that clinical MR images possessing a sparse representation in some transform domain can be
accurately reconstructed even when sampled at rates well below the Nyquist limit by casting the recovery as
a convex ℓ1-minimization problem. While ℓ1-based techniques offer a sizeable advantage over Nyquist-limited
methods, they nonetheless require a modest degree of over-sampling above the true theoretical minimum sampling
rate in order to guarantee the achievability of exact reconstruction. In this work, we present a reconstruction
model based on homotopic approximation of the ℓ0 quasi-norm and discuss the ability of this technique to
reconstruct undersampled MR images at rates even lower than are achievable than with ℓ1-minimization and
arbitrarily close to the true minimum sampling rate. A semi-implicit numerical solver is presented for efficient
numerical computation of the reconstruction process and several examples depicting the capability for accurate
MRI reconstructions from highly-undersampled K-space data are presented.