SCI Publications

In this chapter, we present a high performance multi-scale 3D image processing framework to exploit the parallel processing power of multiple graphic processing units (Multi-GPUs) for medical image analysis. We developed GPU algorithms and data structures that can be applied to a wide range of 3D image processing applications and efficiently exploit the computational power and massive bandwidth offered by modern GPUs. Our framework helps scientists solve computationally intensive problems which previously required super computing power. To demonstrate the effectiveness of our framework and to compare to existing techniques, we focus our discussions on atlas construction - the application of understanding the development of the brain and the progression of brain diseases.

3D iterative CT reconstruction is an active research area in medical imaging. Compared with analytic reconstruction methods such as FDK, iterative methods may provide better reconstruction results for incomplete and noisy projection data. The simultaneous algebraic reconstruction technique (SART), one of the most popular iterative reconstruction methods, is applied in the cone-beam geometry for highresolution reconstruction, with the help of graphics hardware (GPU) and total variation (TV) regularization. GPU greatly improves the efficiency of SART, which is computationally intense for CPU, and thus makes it suitable for clinical applications. TV regularization reduces the effects of noise and helps the convergence of SART for noisy data. Experimental results for both synthetic and real data are provided to evaluate the accuracy and efficiency of the proposed framework.

Keywords: Cone-beam CT, iterative reconstruction, SART, GPU, TV regularization

Unbiased diffeomorphic atlas construction has proven to be a powerful technique for medical image analysis, particularly in brain imaging. The method operates on a large set of images, mapping them all into a common coordinate system, and creating an unbiased common template for studying intra-population variability and interpopulation differences. The technique has also proven effective in tissue and object segmentation via registration of anatomical labels. However, a major barrier to the use of this approach is its high computational cost. Especially with the increasing number of inputs and data size, it becomes impractical even with a fully optimized implementation on CPUs. Fortunately, the highly element-wise independence of the problem makes it well suited for parallel processing. This paper presents an efficient implementation of unbiased diffeomorphic atlas construction on the new parallel processing architecture based on Multi-Graphics Processing Units (Multi-GPUs). Our results show that the GPU implementation gives a substantial performance gain on the order of twenty to sixty times faster than a single CPU and provides an inexpensive alternative to large distributed-memory CPU clusters.

Efficient sorting is a key requirement for many computer science algorithms. Acceleration of existing techniques as well as developing new sorting approaches is crucial for many realtime graphics scenarios, database systems, and numerical simulations to name just a few. It is one of the most fundamental operations to organize and filter the ever growing massive amounts of data gathered on a daily basis. While optimal sorting models for serial execution on a single processor exist, efficient parallel sorting remains a challenge. In this paper we present a hardware-optimized parallel implementation of the radix sort algorithm that results in a significant speed up over existing sorting implementations. We outperform all known GPU based sorting systems by about a factor of two and eliminate restrictions on the sorting key space. This makes our algorithm not only the fastest, but also the first general GPU sorting solution.

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.

Understanding the neural circuitry of the retina requires us to map the connectivity of individual neurons in large neuronal tissue sections and analyze signal communication across processes from the electron microscopy images. One of the major bottlenecks in the critical path is the image mosaicing process where 2D slices are assembled from scanned microscopy image tiles. The problem of assembling the tiles is computationally non-trivial because of distortion of the specimen in the electron microscope due to heat and overlap between the scanned tiles. The complexity of the calculation arises from the massive size of the dataset and mathematical calculations required to calculate value of each pixel of the mosaic. We propose to use texture memory lookups to speedup the access to image tiles and data parallel computing enabled by the GPUs to accelerate this process. The proposed method results in noticeable improvements in speed of computation compared to other methods. Index Terms—Serial-section TEM, image mosaicing, GPGPU, CUDA, texture.

Keywords: crcns, electron microscopy, mosaics, retina, eye