3D computed tomography has been extensively studied and widely used in modern society.
Although most manufacturers choose the filtered backprojection algorithm (FBP) for its accuracy
and efficiency, iterative reconstruction methods have a significant potential to provide superior
performance for incomplete, noisy projection data. However, iterative methods have a high
computational cost, which hinders their practical use. Furthermore, regularization is usually
required to reduce the effects of noise. In this paper, we analyze the use of the Simultaneous
Algebraic Reconstruction Technique (SART) with total variation (TV) regularization.
Additionally, graphics hardware is utilized to increase the speed of SART. NVIDIA's GPU and
Compute Unified Device Architecture (CUDA) comprise the core of our computational platform.
GPU implementation details, including ray-based forward projection and voxel-based
backprojection are illustrated. Experimental results for high-resolution synthetic and real data are
provided to demonstrate the accuracy and efficiency of the proposed framework.
Computed tomography (CT) has been extensively studied and widely used for a variety of medical
applications. The reconstruction of 3D images from a projection series is an important aspect
of the modality. Reconstruction by filtered backprojection (FBP) is used by most manufacturers
because of speed, ease of implementation, and relatively few parameters. Iterative reconstruction
methods have a significant potential to provide superior performance with incomplete or
noisy data, or with less than ideal geometries, such as cone-beam systems. However, iterative
methods have a high computational cost, and regularization is usually required to reduce the
effects of noise. The simultaneous algebraic reconstruction technique (SART) is studied in this
paper, where the Feldkamp method (FDK) for filtered back projection is used as an initialization
for iterative SART. Additionally, graphics hardware is utilized to increase the speed of SART
implementation. Nvidia processors and compute unified device architecture (CUDA) form the
platform for GPU computation. Total variation (TV) minimization is applied for the regularization
of SART results. Preliminary results of SART on 3-D Shepp-Logan phantom using using
TV regularization and GPU computation are presented in this paper. Potential improvements
of the proposed framework are also discussed.
Mobile X-ray imagery is an omnipresent tool in conventional musculoskeletal and soft tissue applications. The next
generation of mobile C-arm systems can provide clinicians of minimally-invasive surgery and pain management
procedures with both real-time high-resolution fluoroscopy and intra-operative CT imaging modalities. In this study, we
research two C-arm CT experimental system configurations and evaluate their imaging capabilities. In a non-destructive
evaluation configuration, the X-ray Tube - Detector assembly is stationary while an imaging object is placed on a
rotating table. In a medical imaging configuration, the C-arm gantry moves around the patient and the table. In our
research setting, we connect the participating devices through a Mobile X-Ray Imaging Environment known as
MOXIE. MOXIE is a set of software applications for internal research at GE Healthcare - Surgery and used to examine
imaging performance of experimental systems. Anthropomorphic phantom volume renderings and orthogonal slices of
reconstructed images are obtained and displayed. The experimental C-arm CT results show CT-like image quality that
may be suitable for interventional procedures, real-time data management, and, therefore, have great potential for
effective use on the clinical floor.
Clinical demands of image-guided procedures present technical challenges in X-ray 1K×1K fluoroscopy and cone-beam
CT on a mobile C-arm. Performance-per-watt and
performance-per-dollar are other major considerations in a search for
an optimal computational platform. Real-time constraints of processing high-resolution fluoroscopic images currently
necessitate the use of highly specialized proprietary image processing hardware, which cannot be easily repurposed for
acceleration of other computing tasks. In our previous studies, we were investigating heterogeneous computing
architectures and suitable hardware/software components to assist in time-critical surgical applications. Through those
studies, it has been shown that Graphics Processing Units (GPUs) can provide outstanding levels of computational
power utilizing the Single Instruction Multiple Data (SIMD) programming model. In the present study, we expand our
research in the domain of real-time processing and continue to explore the feasibility of GPU acceleration for both
fluoroscopic and tomographic imaging. Current emphasis is being placed on applicability of NVIDIA's novel Tesla
computing solutions and Compute Unified Device Architecture (CUDA). The results of this pilot project comprise the
Cg/OpenGL and CUDA algorithm implementations, benchmark evaluations, and examples of processing image data
acquired with use of anthropomorphic phantoms.
The design of mobile X-ray C-arm equipment with image tomography and surgical guidance capabilities involves the
retrieval of repeatable gantry positioning in three-dimensional space. Geometry misrepresentations can cause
degradation of the reconstruction results with the appearance of blurred edges, image artifacts, and even false structures.
It may also amplify surgical instrument tracking errors leading to improper implant placement. In our prior publications
we have proposed a C-arm 3D positioner calibration method comprising separate intrinsic and extrinsic geometry
calibration steps. Following this approach, in the present paper, we extend the intrinsic geometry calibration of C-gantry
beyond angular positions in the orbital plane into angular positions on a unit sphere of isocentric rotation. Our method
makes deployment of markerless interventional tool guidance with use of high-resolution fluoro images and
electromagnetic tracking feasible at any angular position of the tube-detector assembly. Variations of the intrinsic
parameters associated with C-arm motion are measured off-line as functions of orbital and lateral angles. The proposed
calibration procedure provides better accuracy, and prevents unnecessary workflow steps for surgical navigation
applications. With a slight modification, the Misalignment phantom, a tool for intrinsic geometry calibration, is also
utilized to obtain an accurate 'image-to-sensor' mapping. We show simulation results, image quality and navigation
accuracy estimates, and feasibility data acquired with the prototype system. The experimental results show the potential
of high-resolution CT imaging (voxel size below 0.5 mm) and confident navigation in an interventional surgery setting
with a mobile C-arm.
The three-dimensional image reconstruction process used in interventional CT imaging is computationally demanding. Implementation on general-purpose computational platforms requires a substantial time, which is undesirable during time-critical surgical and minimally invasive procedures. Field Programmable Gate Arrays (FPGA)s and Graphics Processing Units (GPU)s have been studied as a platform to accelerate 3-D imaging. FPGA and GPU devices offer a reprogrammable hardware architecture, configurable for pipelining and high levels of parallel processing to increase computational throughput, as well as the benefits of being off-the-shelf and effective 'performance-to-watt' solutions. The main focus of this paper is on the backprojection step of the image reconstruction process, since it is the most computationally intensive part. Using the popular Feldkamp-Davis-Kress (FDK) cone-beam algorithm, our studies indicate the entire 2563 image reconstruction process can be accelerated to real or near real-time (i.e. immediately after a finished scan of 15-30 seconds duration) on a mobile X-ray C-arm system using available resources on built-in FPGA board. High resolution 5123 image backprojection can be also accomplished within the same scanning time on a high-end GPU board comprising up to 128 streaming processors.
Design of C-arm equipment with 3D imaging capabilitys involves retrieval of repeatable gantry positioning information along the acquisition trajectory. Inaccurate retrieval or improper use of positioning information may cause degradation of the reconstruction results, appearance of image artifacts, or indicate false structures. The geometry misrepresentation can also lead to the errors in relative pose assessment of anatomy-of-interest and interventional tools. Comprehensive C gantry calibration with an extended set of misalignment and motion parameters suffers from ambiguity caused by parameter cross-correlation and significant computational complexity. We deploy the concept of a waterfall calibration that comprises sequential intrinsic and extrinsic geometry calibration delineation steps. Following the image-based framework, the first step in our method is intrinsic calibration that deals with delineation of geometry of the X-ray tube-Detector assembly. Extrinsic parameters define motion of the C-arm assembly in 3D space and relate the Camera and World coordinate systems. We formulate both intrinsic and extrinsic calibration problems in vectorized form with total variation constraints. The proposed method has been verified by numerical design and validated by experimental studies. Sequential delineation of intrinsic and extrinsic geometries has demonstrated very efficient performance. The method eliminates the cross-correlation between cone-beam projection parameters, provides significantly better accuracy and computational speed, simplifies the structures of calibration targets used, and avoids the unnecessary workflow and image processing steps. It appears to be adequate for quality and cost derivations in an interventional surgery settings using a mobile C-arm.
CT imaging in interventional and minimally-invasive surgery
requires high-performance computing solutions that meet
operational room demands, healthcare business requirements, and
the constraints of a mobile C-arm system. The computational
requirements of clinical procedures using CT-like data are
increasing rapidly, mainly due to the need for rapid access to
medical imagery during critical surgical procedures. The highly parallel nature of Radon transform and CT algorithms enables
embedded computing solutions utilizing a parallel processing
architecture to realize a significant gain of computational
intensity with comparable hardware and program coding/testing
expenses. In this paper, using a sample 2D and 3D CT problem, we
explore the programming challenges and the potential benefits of
embedded computing using commodity hardware components. The
accuracy and performance results obtained on three computational
platforms: a single CPU, a single GPU, and a solution based on FPGA technology have been analyzed. We have shown that
hardware-accelerated CT image reconstruction can be achieved with
similar levels of noise and clarity of feature when compared to
program execution on a CPU, but gaining a performance increase at
one or more orders of magnitude faster. 3D cone-beam or helical CT
reconstruction and a variety of volumetric image processing
applications will benefit from similar accelerations.