Tunable neural networks for CT image formation.

Optimization of CT image quality typically involves balancing variance and bias. In traditional filtered back-projection, this trade-off is controlled by the filter cutoff frequency. In model-based iterative reconstruction, the regularization strength parameter often serves the same function. Deep neural networks (DNNs) typically do not provide this tunable control over output image properties. Models are often trained to minimize the expected mean squared error, which penalizes both variance and bias in image outputs but does not offer any control over the trade-off between the two. We propose a method for controlling the output image properties of neural networks with a new loss function called weighted covariance and bias (WCB). Our proposed method uses multiple noise realizations of the input images during training to allow for separate weighting matrices for the variance and bias penalty terms. Moreover, we show that tuning these weights enables targeted penalization of specific image features with spatial frequency domain penalties. To evaluate our method, we present a simulation study using digital anthropomorphic phantoms, physical simulation of CT measurements, and image formation with various algorithms. We show that the WCB loss function offers a greater degree of control over trade-offs between variance and bias, whereas mean-squared error provides only one specific image quality configuration. We also show that WCB can be used to control specific image properties including variance, bias, spatial resolution, and the noise correlation of neural network outputs. Finally, we present a method to optimize the proposed weights for a spiculated lung nodule shape discrimination task. Our results demonstrate this new image quality can control the image properties of DNN outputs and optimize image quality for task-specific performance.

Flat-panel conebeam CT in the clinic: history and current state.

Research into conebeam CT concepts began as soon as the first clinical single-slice CT scanner was conceived. Early implementations of conebeam CT in the 1980s focused on high-contrast applications where concurrent high resolution ( < 200 µ m ), for visualization of small contrast-filled vessels, bones, or teeth, was an imaging requirement that could not be met by the contemporaneous CT scanners. However, the use of nonlinear imagers, e.g., x-ray image intensifiers, limited the clinical utility of the earliest diagnostic conebeam CT systems. The development of consumer-electronics large-area displays provided a technical foundation that was leveraged in the 1990s to first produce large-area digital x-ray detectors for use in radiography and then compact flat panels suitable for high-resolution and high-frame-rate conebeam CT. In this review, we show the concurrent evolution of digital flat panel (DFP) technology and clinical conebeam CT. We give a brief summary of conebeam CT reconstruction, followed by a brief review of the correction approaches for DFP-specific artifacts. The historical development and current status of flat-panel conebeam CT in four clinical areas-breast, fixed C-arm, image-guided radiation therapy, and extremity/head-is presented. Advances in DFP technology over the past two decades have led to improved visualization of high-contrast, high-resolution clinical tasks, and image quality now approaches the soft-tissue contrast resolution that is the standard in clinical CT. Future technical developments in DFPs will enable an even broader range of clinical applications; research in the arena of flat-panel CT shows no signs of slowing down.

Known-component 3D image reconstruction for improved intraoperative imaging in spine surgery: A clinical pilot study.

PURPOSE: Intraoperative imaging plays an increased role in support of surgical guidance and quality assurance for interventional approaches. However, image quality sufficient to detect complications and provide quantitative assessment of the surgical product is often confounded by image noise and artifacts. In this work, we translated a three-dimensional model-based image reconstruction (referred to as "Known-Component Reconstruction," KC-Recon) for the first time to clinical studies with the aim of resolving both limitations. METHODS: KC-Recon builds upon a penalized weighted least-squares (PWLS) method by incorporating models of surgical instrumentation ("known components") within a joint image registration-reconstruction process to improve image quality. Under IRB approval, a clinical pilot study was conducted with 17 spine surgery patients imaged under informed consent using the O-arm cone-beam CT system (Medtronic, Littleton MA) before and after spinal instrumentation. Volumetric images were generated for each patient using KC-Recon in comparison to conventional filtered backprojection (FBP). Imaging performance prior to instrumentation ("preinstrumentation") was evaluated in terms of soft-tissue contrast-to-noise ratio (CNR) and spatial resolution. The quality of images obtained after the instrumentation ("postinstrumentation") was assessed by quantifying the magnitude of metal artifacts (blooming and streaks) arising from pedicle screws. The potential low-dose advantages of the algorithm were tested by simulating low-dose data (down to one-tenth of the dose of standard protocols) from images acquired at normal dose. RESULTS: Preinstrumentation images (at normal clinical dose and matched resolution) exhibited an average 24.0% increase in soft-tissue CNR with KC-Recon compared to FBP (N = 16, P = 0.02), improving visualization of paraspinal muscles, major vessels, and other soft-tissues about the spine and abdomen. For a total of 72 screws in postinstrumentation images, KC-Recon yielded a significant reduction in metal artifacts: 66.3% reduction in overestimation of screw shaft width due to blooming (P < 0.0001) and reduction in streaks at the screw tip (65.8% increase in attenuation accuracy, P < 0.0001), enabling clearer depiction of the screw within the pedicle and vertebral body for an assessment of breach. Depending on the imaging task, dose reduction up to an order of magnitude appeared feasible while maintaining soft-tissue visibility and metal artifact reduction. CONCLUSIONS: KC-Recon offers a promising means to improve visualization in the presence of surgical instrumentation and reduce patient dose in image-guided procedures. The improved soft-tissue visibility could facilitate the use of cone-beam CT to soft-tissue surgeries, and the ability to precisely quantify and visualize instrument placement could provide a valuable check against complications in the operating room (cf., postoperative CT).

Model-based Multi-material Decomposition using Spatial-Spectral CT Filters.

Spectral CT with multiple contrast agents has been enabled by energy-discriminating detectors with multiple spectral channels. We propose a new approach that uses spatial-spectral filters to provide multiple beamlets with different incident spectra for spectral channels based on "source-side" control. Since these spatial-spectral filters yield spectral channels that are sparse, we adopt model-based material decomposition to directly reconstruct material densities from projection data. Simulation studies in three-and four-material decomposition experiments show the underlying feasibility of the spatial-spectral filtering technique. This methodology has the potential to facilitate imaging of multiple contrast agents simultaneously with relatively simple hardware, or to improve spectral CT performance via combination with other established spectral CT methods for additional control and flexibility.

High-Fidelity Modeling of Detector Lag and Gantry Motion in CT Reconstruction.

Detector lag and gantry motion during x-ray exposure and integration both result in azimuthal blurring in CT reconstructions. These effects can degrade image quality both for high-resolution features as well as low-contrast details. In this work we consider a forward model for model-based iterative reconstruction (MBIR) that is sufficiently general to accommodate both of these physical effects. We integrate this forward model in a penalized, weighted, nonlinear least-square style objective function for joint reconstruction and correction of these blur effects. We show that modeling detector lag can reduce/remove the characteristic lag artifacts in head imaging in both a simulation study and physical experiments. Similarly, we show that azimuthal blur ordinarily introduced by gantry motion can be mitigated with proper reconstruction models. In particular, we find the largest image quality improvement at the periphery of the field-of-view where gantry motion artifacts are most pronounced. These experiments illustrate the generality of the underlying forward model, suggesting the potential application in modeling a number of physical effects that are traditionally ignored or mitigated through pre-corrections to measurement data.

Dynamic fluence field modulation for miscentered patients in computed tomography.

Traditional CT image acquisition uses bowtie filters to reduce dose, x-ray scatter, and detector dynamic range requirements. However, accurate patient centering within the bore of the CT scanner takes time and is often difficult to achieve precisely. Patient miscentering combined with a static bowtie filter can result in significant increases in dose, reconstruction noise, and CT number variations, and consequently raise overall exposure requirements. Approaches to estimate the patient position from scout scans and perform dynamic spatial beam filtration during acquisition are developed and applied in physical experiments on a CT test bench using different beam filtration strategies. While various dynamic beam modulation strategies have been developed, we focus on two approaches: (1) a simple approach using attenuation-based beam modulation using a translating bowtie filter and (2) dynamic beam modulation using multiple aperture devices (MADs)-an emerging beam filtration strategy based on binary filtration of the x-ray beam using variable width slits in a high-density beam blocker. Improved dose utilization and more consistent image performance with respect to an unmodulated baseline (static filter) are demonstrated for miscentered objects and dynamic beam filtration in physical experiments. For a homogeneous object miscentered by 4 cm, the dynamic filter reduced the maximum regional noise and dose penalties (compared with a centered object) from 173% to 16% and 42% to 14%, respectively, for a traditional bowtie, 29% to 8% and 24% to 15%, respectively, for a single MAD, and 275% to 11% and 56% to 18%, respectively, for a dual-MAD filter. The proposed methodology has the potential to relax patient centering requirements within the scanner, reduce setup time, and facilitate additional CT dose reduction.

Task-driven optimization of CT tube current modulation and regularization in model-based iterative reconstruction.

Tube current modulation (TCM) is routinely adopted on diagnostic CT scanners for dose reduction. Conventional TCM strategies are generally designed for filtered-backprojection (FBP) reconstruction to satisfy simple image quality requirements based on noise. This work investigates TCM designs for model-based iterative reconstruction (MBIR) to achieve optimal imaging performance as determined by a task-based image quality metric. Additionally, regularization is an important aspect of MBIR that is jointly optimized with TCM, and includes both the regularization strength that controls overall smoothness as well as directional weights that permits control of the isotropy/anisotropy of the local noise and resolution properties. Initial investigations focus on a known imaging task at a single location in the image volume. The framework adopts Fourier and analytical approximations for fast estimation of the local noise power spectrum (NPS) and modulation transfer function (MTF)-each carrying dependencies on TCM and regularization. For the single location optimization, the local detectability index (d') of the specific task was directly adopted as the objective function. A covariance matrix adaptation evolution strategy (CMA-ES) algorithm was employed to identify the optimal combination of imaging parameters. Evaluations of both conventional and task-driven approaches were performed in an abdomen phantom for a mid-frequency discrimination task in the kidney. Among the conventional strategies, the TCM pattern optimal for FBP using a minimum variance criterion yielded a worse task-based performance compared to an unmodulated strategy when applied to MBIR. Moreover, task-driven TCM designs for MBIR were found to have the opposite behavior from conventional designs for FBP, with greater fluence assigned to the less attenuating views of the abdomen and less fluence to the more attenuating lateral views. Such TCM patterns exaggerate the intrinsic anisotropy of the MTF and NPS as a result of the data weighting in MBIR. Directional penalty design was found to reinforce the same trend. The task-driven approaches outperform conventional approaches, with the maximum improvement in d' of 13% given by the joint optimization of TCM and regularization. This work demonstrates that the TCM optimal for MBIR is distinct from conventional strategies proposed for FBP reconstruction and strategies optimal for FBP are suboptimal and may even reduce performance when applied to MBIR. The task-driven imaging framework offers a promising approach for optimizing acquisition and reconstruction for MBIR that can improve imaging performance and/or dose utilization beyond conventional imaging strategies.

Design of dual multiple aperture devices for dynamical fluence field modulated CT.

A Multiple Aperture Device (MAD) is a novel x-ray beam modulator that uses binary filtration on a fine scale to spatially modulate an x-ray beam. Using two MADs in series enables a large variety of fluence profiles by shifting the MADS relative to each other. This work details the design and control of dual MADs for a specific class of desired fluence patterns. Specifically, models of MAD operation are integrated into a best fit objective followed by CMA-ES optimization. To illustrate this framework we demonstrate the design process for an abdominal phantom with the goal of uniform detected signal. Achievable fluence profiles show good agreement with target fluence profiles, and the ability to flatten projections when a phantom is scanned is demonstrated. Simulated data reconstruction using traditional tube current modulation (TCM) and MAD filtering with TCM are investigated with the dual MAD system demonstrating more uniformity in noise and illustrating the potential for dose reduction under a maximum noise level constraint.

Robust 3D-2D image registration: application to spine interventions and vertebral labeling in the presence of anatomical deformation.

We present a framework for robustly estimating registration between a 3D volume image and a 2D projection image and evaluate its precision and robustness in spine interventions for vertebral localization in the presence of anatomical deformation. The framework employs a normalized gradient information similarity metric and multi-start covariance matrix adaptation evolution strategy optimization with local-restarts, which provided improved robustness against deformation and content mismatch. The parallelized implementation allowed orders-of-magnitude acceleration in computation time and improved the robustness of registration via multi-start global optimization. Experiments involved a cadaver specimen and two CT datasets (supine and prone) and 36 C-arm fluoroscopy images acquired with the specimen in four positions (supine, prone, supine with lordosis, prone with kyphosis), three regions (thoracic, abdominal, and lumbar), and three levels of geometric magnification (1.7, 2.0, 2.4). Registration accuracy was evaluated in terms of projection distance error (PDE) between the estimated and true target points in the projection image, including 14 400 random trials (200 trials on the 72 registration scenarios) with initialization error up to ±200 mm and ±10°. The resulting median PDE was better than 0.1 mm in all cases, depending somewhat on the resolution of input CT and fluoroscopy images. The cadaver experiments illustrated the tradeoff between robustness and computation time, yielding a success rate of 99.993% in vertebral labeling (with 'success' defined as PDE <5 mm) using 1,718 664 ± 96 582 function evaluations computed in 54.0 ± 3.5 s on a mid-range GPU (nVidia, GeForce GTX690). Parameters yielding a faster search (e.g., fewer multi-starts) reduced robustness under conditions of large deformation and poor initialization (99.535% success for the same data registered in 13.1 s), but given good initialization (e.g., ±5 mm, assuming a robust initial run) the same registration could be solved with 99.993% success in 6.3 s. The ability to register CT to fluoroscopy in a manner robust to patient deformation could be valuable in applications such as radiation therapy, interventional radiology, and an assistant to target localization (e.g., vertebral labeling) in image-guided spine surgery.

Cascaded systems analysis of noise and detectability in dual-energy cone-beam CT.

PURPOSE: Dual-energy computed tomography and dual-energy cone-beam computed tomography (DE-CBCT) are promising modalities for applications ranging from vascular to breast, renal, hepatic, and musculoskeletal imaging. Accordingly, the optimization of imaging techniques for such applications would benefit significantly from a general theoretical description of image quality that properly incorporates factors of acquisition, reconstruction, and tissue decomposition in DE tomography. This work reports a cascaded systems analysis model that includes the Poisson statistics of x rays (quantum noise), detector model (flat-panel detectors), anatomical background, image reconstruction (filtered backprojection), DE decomposition (weighted subtraction), and simple observer models to yield a task-based framework for DE technique optimization. METHODS: The theoretical framework extends previous modeling of DE projection radiography and CBCT. Signal and noise transfer characteristics are propagated through physical and mathematical stages of image formation and reconstruction. Dual-energy decomposition was modeled according to weighted subtraction of low- and high-energy images to yield the 3D DE noise-power spectrum (NPS) and noise-equivalent quanta (NEQ), which, in combination with observer models and the imaging task, yields the dual-energy detectability index (d(')). Model calculations were validated with NPS and NEQ measurements from an experimental imaging bench simulating the geometry of a dedicated musculoskeletal extremities scanner. Imaging techniques, including kVp pair and dose allocation, were optimized using d(') as an objective function for three example imaging tasks: (1) kidney stone discrimination; (2) iodine vs bone in a uniform, soft-tissue background; and (3) soft tissue tumor detection on power-law anatomical background. RESULTS: Theoretical calculations of DE NPS and NEQ demonstrated good agreement with experimental measurements over a broad range of imaging conditions. Optimization results suggest a lower fraction of total dose imparted by the low-energy acquisition, a finding consistent with previous literature. The selection of optimal kVp pair reveals the combined effect of both quantum noise and contrast in the kidney stone discrimination and soft-tissue tumor detection tasks, whereas the K-edge effect of iodine was the dominant factor in determining kVp pairs in the iodine vs bone task. The soft-tissue tumor task illustrated the benefit of dual-energy imaging in eliminating anatomical background noise and improving detectability beyond that achievable by single-energy scans. CONCLUSIONS: This work established a task-based theoretical framework that is predictive of DE image quality. The model can be utilized in optimizing a broad range of parameters in image acquisition, reconstruction, and decomposition, providing a useful tool for maximizing DE-CBCT image quality and reducing dose.