Hannes G. Hofmann

CONRAD--a software framework for cone-beam imaging in radiology.

PURPOSE In the community of x-ray imaging, there is a multitude of tools and applications that are used in scientific practice. Many of these tools are proprietary and can only be used within a certain lab. Often the same algorithm is implemented multiple times by different groups in order to enable comparison. In an effort to tackle this problem, the authors created CONRAD, a software framework that provides many of the tools that are required to simulate basic processes in x-ray imaging and perform image reconstruction with consideration of nonlinear physical effects. METHODS CONRAD is a Java-based state-of-the-art software platform with extensive documentation. It is based on platform-independent technologies. Special libraries offer access to hardware acceleration such as OpenCL. There is an easy-to-use interface for parallel processing. The software package includes different simulation tools that are able to generate up to 4D projection and volume data and respective vector motion fields. Well known reconstruction algorithms such as FBP, DBP, and ART are included. All algorithms in the package are referenced to a scientific source. RESULTS A total of 13 different phantoms and 30 processing steps have already been integrated into the platform at the time of writing. The platform comprises 74.000 nonblank lines of code out of which 19% are used for documentation. The software package is available for download at http://conrad.stanford.edu. To demonstrate the use of the package, the authors reconstructed images from two different scanners, a table top system and a clinical C-arm system. Runtimes were evaluated using the RabbitCT platform and demonstrate state-of-the-art runtimes with 2.5 s for the 256 problem size and 12.4 s for the 512 problem size. CONCLUSIONS As a common software framework, CONRAD enables the medical physics community to share algorithms and develop new ideas. In particular this offers new opportunities for scientific collaboration and quantitative performance comparison between the methods of different groups.

GPU-accelerated SART reconstruction using the CUDA programming environment

The Common Unified Device Architecture (CUDA) introduced in 2007 by NVIDIA is a recent programming model making use of the unified shader design of the most recent graphics processing units (GPUs). The programming interface allows algorithm implementation using standard C language along with a few extensions without any knowledge about graphics programming using OpenGL, DirectX, and shading languages. We apply this novel technology to the Simultaneous Algebraic Reconstruction Technique (SART), which is an advanced iterative image reconstruction method in cone-beam CT. So far, the computational complexity of this algorithm has prohibited its use in most medical applications. However, since todays GPUs provide a high level of parallelism and are highly cost-efficient processors, they are predestinated for performing the iterative reconstruction according to medical requirements. In this paper we present an efficient implementation of the most time-consuming parts of the iterative reconstruction algorithm: forward- and back-projection. We also explain the required strategy to parallelize the algorithm for the CUDA 1.1 and CUDA 2.0 architecture. Furthermore, our implementation introduces an acceleration technique for the reconstruction compared to a standard SART implementation on the GPU using CUDA. Thus, we present an implementation that can be used in a time-critical clinical environment. Finally, we compare our results to the current applications on multi-core workstations, with respect to both reconstruction speed and (dis-)advantages. Our implementation exhibits a speed-up of more than 64 compared to a state-of-the-art CPU using hardware-accelerated texture interpolation.

Technical Note: RabbitCT―an open platform for benchmarking 3D cone-beam reconstruction algorithmsa)

PURPOSE Fast 3D cone beam reconstruction is mandatory for many clinical workflows. For that reason, researchers and industry work hard on hardware-optimized 3D reconstruction. Backprojection is a major component of many reconstruction algorithms that require a projection of each voxel onto the projection data, including data interpolation, before updating the voxel value. This step is the bottleneck of most reconstruction algorithms and the focus of optimization in recent publications. A crucial limitation, however, of these publications is that the presented results are not comparable to each other. This is mainly due to variations in data acquisitions, preprocessing, and chosen geometries and the lack of a common publicly available test dataset. The authors provide such a standardized dataset that allows for substantial comparison of hardware accelerated backprojection methods. METHODS They developed an open platform RabbitCT (www.rabbitCT.com) for worldwide comparison in backprojection performance and ranking on different architectures using a specific high resolution C-arm CT dataset of a rabbit. This includes a sophisticated benchmark interface, a prototype implementation in C++, and image quality measures. RESULTS At the time of writing, six backprojection implementations are already listed on the website. Optimizations include multithreading using Intel threading building blocks and OpenMP, vectorization using SSE, and computation on the GPU using CUDA 2.0. CONCLUSIONS There is a need for objectively comparing backprojection implementations for reconstruction algorithms. RabbitCT aims to provide a solution to this problem by offering an open platform with fair chances for all participants. The authors are looking forward to a growing community and await feedback regarding future evaluations of novel software- and hardware-based acceleration schemes.

A Comparative Error Analysis of Current Time-of-Flight Sensors

Time-of-flight (ToF) cameras suffer from systematic errors, which can be an issue in many application scenarios. In this paper, we investigate the error characteristics of eight different ToF cameras. Our survey covers both well established and recent cameras including the Microsoft Kinect V2. We present up to six experiments for each camera to quantify different types of errors. For each experiment, we outline the basic setup, present comparable data for each camera, and discuss the respective results. The results discussed in this paper enable the community to make appropriate decisions in choosing the best matching camera for a certain application. This work also lays the foundation for a framework to benchmark future ToF cameras. Furthermore, our results demonstrate the necessity for correcting characteristic measurement errors. We believe that the presented findings will allow 1) the development of novel correction methods for specific errors and 2) the development of general data processing algorithms that are able to robustly operate on a wider range of cameras and scenes.

Fast simulation of x-ray projections of spline-based surfaces using an append buffer

Many scientists in the field of x-ray imaging rely on the simulation of x-ray images. As the phantom models become more and more realistic, their projection requires high computational effort. Since x-ray images are based on transmission, many standard graphics acceleration algorithms cannot be applied to this task. However, if adapted properly, the simulation speed can be increased dramatically using state-of-the-art graphics hardware. A custom graphics pipeline that simulates transmission projections for tomographic reconstruction was implemented based on moving spline surface models. All steps from tessellation of the splines, projection onto the detector and drawing are implemented in OpenCL. We introduced a special append buffer for increased performance in order to store the intersections with the scene for every ray. Intersections are then sorted and resolved to materials. Lastly, an absorption model is evaluated to yield an absorption value for each projection pixel. Projection of a moving spline structure is fast and accurate. Projections of size 640 × 480 can be generated within 254 ms. Reconstructions using the projections show errors below 1 HU with a sharp reconstruction kernel. Traditional GPU-based acceleration schemes are not suitable for our reconstruction task. Even in the absence of noise, they result in errors up to 9 HU on average, although projection images appear to be correct under visual examination. Projections generated with our new method are suitable for the validation of novel CT reconstruction algorithms. For complex simulations, such as the evaluation of motion-compensated reconstruction algorithms, this kind of x-ray simulation will reduce the computation time dramatically.

Three-dimensional anisotropic adaptive filtering of projection data for noise reduction in cone beam CT

PURPOSE The combination of quickly rotating C-arm gantry with digital flat panel has enabled the acquisition of three-dimensional data (3D) in the interventional suite. However, image quality is still somewhat limited since the hardware has not been optimized for CT imaging. Adaptive anisotropic filtering has the ability to improve image quality by reducing the noise level and therewith the radiation dose without introducing noticeable blurring. By applying the filtering prior to 3D reconstruction, noise-induced streak artifacts are reduced as compared to processing in the image domain. METHODS 3D anisotropic adaptive filtering was used to process an ensemble of 2D x-ray views acquired along a circular trajectory around an object. After arranging the input data into a 3D space (2D projections + angle), the orientation of structures was estimated using a set of differently oriented filters. The resulting tensor representation of local orientation was utilized to control the anisotropic filtering. Low-pass filtering is applied only along structures to maintain high spatial frequency components perpendicular to these. The evaluation of the proposed algorithm includes numerical simulations, phantom experiments, and in-vivo data which were acquired using an AXIOM Artis dTA C-arm system (Siemens AG, Healthcare Sector, Forchheim, Germany). Spatial resolution and noise levels were compared with and without adaptive filtering. A human observer study was carried out to evaluate low-contrast detectability. RESULTS The adaptive anisotropic filtering algorithm was found to significantly improve low-contrast detectability by reducing the noise level by half (reduction of the standard deviation in certain areas from 74 to 30 HU). Virtually no degradation of high contrast spatial resolution was observed in the modulation transfer function (MTF) analysis. Although the algorithm is computationally intensive, hardware acceleration using Nvidias CUDA Interface provided an 8.9-fold speed-up of the processing (from 1336 to 150 s). CONCLUSIONS Adaptive anisotropic filtering has the potential to substantially improve image quality and/or reduce the radiation dose required for obtaining 3D image data using cone beam CT.

Pushing the limits for medical image reconstruction on recent standard multicore processors

Volume reconstruction by backprojection is the computational bottleneck in many interventional clinical computed tomography (CT) applications. Today vendors in this field replace special purpose hardware accelerators with standard hardware such as multicore chips and GPGPUs. Medical imaging algorithms are on the verge of employing high-performance computing (HPC) technology, and are therefore an interesting new candidate for optimization. This paper presents low-level optimizations for the backprojection algorithm, guided by a thorough performance analysis on four generations of Intel multicore processors (Harpertown, Westmere, Westmere EX, and Sandy Bridge). We choose the RabbitCT benchmark, a standardized testcase well supported in industry, to ensure transparent and comparable results. Our aim is to provide not only the fastest possible implementation but also compare with performance models and hardware counter data in order to fully understand the results. We separate the influence of algorithmic optimizations, parallelization, SIMD vectorization, and microarchitectural issues and pinpoint problems with current SIMD instruction set extensions on standard CPUs (SSE, AVX). The use of assembly language is mandatory for best performance. Finally, we compare our results to the best GPGPU implementations available for this open competition benchmark.

Real-Time Range Imaging in Health Care: A Survey

The recent availability of dynamic, dense, and low-cost range imaging has gained widespread interest in health care. It opens up new opportunities and has an increasing impact on both research and commercial activities. This chapter presents a state-of-the-art survey on the integration of modern range imaging sensors into medical applications. The scope is to identify promising applications and methods, and to provide an overview of recent developments in this rapidly evolving domain. The survey covers a broad range of topics, including guidance in computer-assisted interventions, operation room monitoring and workflow analysis, touch-less interaction and on-patient visualization, as well as prevention and support in elderly care and rehabilitation. We put emphasis on dynamic and interactive tasks where real-time and dense 3-D imaging forms the key aspect. While considering different range imaging modalities that fulfill these requirements, we particularly investigate the impact of Time-of-Flight imaging in this domain. Eventually, we discuss practical demands and limitations, and open research issues and challenges that are of fundamental importance for the progression of the field.

Evaluation of state-of-the-art hardware architectures for fast cone-beam CT reconstruction

We present an evaluation of state-of-the-art computer hardware architectures for implementing the FDK method, which solves the 3-D image reconstruction task in cone-beam computed tomography (CT). The computational complexity of the FDK method prohibits its use for many clinical applications unless appropriate hardware acceleration is employed. Todays most powerful hardware architectures for high-performance computing applications are based on standard multi-core processors, off-the-shelf graphics boards, the Cell Broadband Engine Architecture (CBEA), or customized accelerator platforms (e.g., FPGA-based computer components). For each hardware platform under consideration, we describe a thoroughly optimized implementation of the most time-consuming parts of the FDK algorithm; the filtering step as well as the subsequent back-projection step. We further explain the required code transformations to parallelize the algorithm for the respective target architecture. We compare both the implementation complexity and the resulting performance of all architectures under consideration using the same two medical datasets which have been acquired using a standard C-arm device. Our optimized back-projection implementations achieve at least a speedup of 6.5 (CBEA, two processors), 22.0 (GPU, single board), and 35.8 (FPGA, 9 chips) compared to a standard workstation equipped with a quad-core processor.

Implementation of the FDK algorithm for cone-beam CT on the cell broadband engine architecture

In most of todays commercially available cone-beam CT scanners, the well known FDK method is used for solving the 3D reconstruction task. The computational complexity of this algorithm prohibits its use for many medical applications without hardware acceleration. The brand-new Cell Broadband Engine Architecture (CBEA) with its high level of parallelism is a cost-efficient processor for performing the FDK reconstruction according to the medical requirements. The programming scheme, however, is quite different to any standard personal computer hardware. In this paper, we present an innovative implementation of the most time-consuming parts of the FDK algorithm: filtering and back-projection. We also explain the required transformations to parallelize the algorithm for the CBEA. Our software framework allows to compute the filtering and back-projection in parallel, making it possible to do an on-the-fly-reconstruction. The achieved results demonstrate that a complete FDK reconstruction is computed with the CBEA in less than seven seconds for a standard clinical scenario. Given the fact that scan times are usually much higher, we conclude that reconstruction is finished right after the end of data acquisition. This enables us to present the reconstructed volume to the physician in real-time, immediately after the last projection image has been acquired by the scanning device.