Denoising is an indispensable step in processing low-dose X-ray fluoroscopic images that requires development of specialized high-quality algorithms able to operate in near real-time. We address this problem with an efficient deep learning approach based on the process-centric view of traditional iterative thresholding methods. We develop a novel trainable patch-based multiscale framework for sparse image representation. In a computationally efficient way, it allows us to accurately reconstruct important image features on multiple levels of decomposition with patch dictionaries of reduced size and complexity. The flexibility of the chosen machine learning approach allows us to tailor the learned basis for preserving important structural information in the image and noticeably minimize the amount of artifacts. Our denoising results obtained with real clinical data demonstrate significant quality improvement and are computed much faster in comparison with the BM3D algorithm.

While initial compressed sensing (CS) recovery techniques operated under the implicit assumption that the sparse domain coefficients are independently distributed, recent results have indicated that integrating a statistical or structural dependence model of sparse domain coefficients into CS enhances recovery. In this paper, we present a method for exploiting empirical dependences among wavelet coefficients during CS recovery using a Bayes least-square Gaussian-scale-mixture model. The proposed model is successfully incorporated into several recent CS algorithms, including reweighted l(1) minimization (RL1), iteratively reweighted least squares, and iterative hard thresholding. Extensive experiments including comparisons with a state-of-the-art model-based CS method demonstrate that the proposed algorithms are highly effective at reducing reconstruction error and/or the number of measurements required for a desired reconstruction quality.

Dynamic magnetic resonance imaging (MRI) is commonly used to observe dynamic physiological changes in tissue or to study organs with mobile structures such as the heart. In order to accurately capture spatiotemporal changes, it is desirable to have dynamic images with high temporal resolution in addition to high spatial resolution. Due to the nature of data acquisition in current MRI systems, there exists a trade-off between temporal and spatial resolution. In this work, we present two methods for improving the spatiotemporal resolution in dynamic MRI using compressive sampling (CS). Experimental results illustrate that the proposed Bayes least squares-Gaussian scale mixtures (BLS-GSM) model-based CS algorithm compares favorably with other state-of-the-art compressive dynamic MRI techniques.

This paper presents a method for improving wavelet-based Compressed Sensing (CS) reconstruction algorithms by exploiting the dependencies among wavelet coefficients. During CS recovery, a simple measure of significance for each wavelet coefficient is calculated using a weighted sum of the (estimated) magnitudes of the wavelet coefficient, its highly correlated neighbors, and parent. This simple measure is incorporated into three CS recovery algorithms, Reweighted L1 minimization algorithms (RL1), Iteratively Reweighted Least Squares (IRLS), and Iterative Hard Thresholding (IHT). Experimental results using one-dimensional signals and images illustrate that the proposed method (i) improves reconstruction quality for a given number of measurements, (ii) requires fewer measurements for a desired reconstruction quality, and (iii) significantly reduces reconstruction time.

Compressed Sensing (CS) theory has gained attention recently as an alternative to the current paradigm of sampling followed by compression. Early CS recovery techniques operated under the implicit assumption that the transform coefficients in the sparsity domain are independently distributed. Recent works, however, demonstrated that exploiting the statistical dependencies between transform coefficients can further improve the recovery performance of CS. In this paper, we propose the use of a Gaussian Scale Mixtures (GSM) model in CS. This model can efficiently exploit the statistical dependencies between wavelet coefficients during CS recovery. The proposed model is incorporated into several recent CS techniques including Reweighted l1minimization (RL1), Iteratively Reweighted Least Squares (IRLS), and Iterative Hard Thresholding (IHT). Experimental results show that the proposed method improves reconstruction quality for a given number of measurements or requires fewer measurements for a desired reconstruction quality.

Remote visualization of volumetric images has gained importance over the past few years in medical and industrial applications. Volume visualization is a computationally intensive process, often requiring hardware acceleration to achieve a real time viewing experience. One remote visualization model that can accomplish this would transmit rendered images from a server, based on viewpoint requests from a client. For constrained server-client bandwidth, an efficient compression scheme is vital for transmitting high quality rendered images. In this paper, we present a new view compensation scheme that utilizes the geometric relationship between viewpoints to exploit the correlation between successive rendered images. The proposed method obviates motion estimation between rendered images, enabling significant reduction to the complexity of a compressor. Additionally, the view compensation scheme, in conjunction with JPEG2000 performs better than AVC, the state of the art video compression standard.

Summary form only given. JPEG2000 Part 2 by Taubman, D.S. and Marcellin, M.W. (2002) supports the use of multi-component transforms (MCT) to decorrelate multi-component image data along the component direction. Three types of point transforms are allowed: linear block transforms, dependency transforms and wavelet transforms. Besides these transforms, Part 2 also allows grouping of arbitrary subsets of components into 'component collections.' Any of the three transform types can be used for each such collection. These Part 2 extensions have been used for compressing 3-D images in many applications such as medical, hyper-spectral and video. It is widely believed that the MCT extensions are applicable only to 3-D data. In this work, we demonstrate their use for compressing N-datasets for any N ges 3. Let (M1, M2,..., MN-2, X, Y) denote the dimensions (size) of an N-dataset. The data can then be interpreted as a collection of 2-D (X, Y) images (or components). The number of such components is T = pik=1N-2Mk. Each component is indexed by an N - 2 dimensional vector: [m1,m2,...,mN-2] where mkisin {1, 2,..., Mk}. To compress an N-D dataset, the T components are processed by N - 2 stages of MCT. Specifically, each stage has T input components and produces T output components, which in turn are used as inputs to the next stage. Consider the kth(k isin {1, 2,..., N - 2}) stage of the MCT which is used to decorrelate the data along the k dimension. The input components are first reordered to form collections. A collection is formed by grouping all components having the same value of the N - 3 dimensional index [m1, m2,...,mk+1,...mN-2]. There are then T/Mkcollections, each containing Mkcomponents. For each collection, a point transform is applied, resulting in decorrelation of the data along the kthdimension. The output components of the final MCT stage (N - 2) are compressed using the 2-D wavelet transform and block coding of JPEG2000. It should be noted that the encoding procedure described above will result in a code-stream that is entirely compliant with JPEG2000 Part 2. Two major applications of multi-dimensional sources are medical images (fMRI, 4-D cardiac and ultrasound) and multi-view video. We have applied the proposed methodology to a 5-D (X, Y, slice, time, trial) fMRI dataset. A PSNR (Peak Signal to Noise Ratio) gain of 8 to 14.9 dB is observed at various bit rates, when 3 stages of MCT (5-D) are used as opposed to the usual single stage of MCT (3-D). The proposed scheme can support scalability in all N dimensions. Additionally, incremental processing can be used to achieve considerable savings in memory usage.

Remote visualization of volumetric images has gained importance over the past few years in medical and industrial applications. Volume visualization is a computationally intensive process, often requiring hardware acceleration to achieve a real time viewing experience. One remote visualization model that can accomplish this would transmit rendered images (with dedicated hardware) from the server based on view-point requests from a client. For a given server-client bandwidth, an efficient compression scheme is vital for transmitting high quality rendered images. In this paper, we present a new compression scheme that utilizes the geometric relationship between view-points to exploit the correlation between successive rendered images. The proposed method performs better than AVC, the state of the art video compression standard. Additionally, our scheme obviates motion estimation between rendered images, enabling significant reduction to the complexity of the encoder.

JPEG2000 Part 2 supports the use of multicomponent transforms (MCTs) to decorrelate multicomponent images along the component direction. Such point transforms can be performed on arbitrary subsets of components, known as "component collections." These Part 2 extensions have been used for compressing 3-D images in applications such as medical imaging and remote sensing. It is widely believed that the MCT extensions are applicable only to 3-D data. In this letter, we demonstrate their use for compressing N-D datasets for any N ges 3.

We present a method for stochastic fiber tract mapping from diffusion tensor MRI (DT-MRI) implemented on graphics hardware. From the simulated fibers we compute a connectivity map that gives an indication of the probability that two points in the dataset are connected by a neuronal fiber path. A Bayesian formulation of the fiber model is given and it is shown that the inversion method can be used to construct plausible connectivity. An implementation of this fiber model on the graphics processing unit (GPU) is presented. Since the fiber paths can be stochastically generated independently of one another, the algorithm is highly parallelizable. This allows us to exploit the data-parallel nature of the GPU fragment processors. We also present a framework for the connectivity computation on the GPU. Our implementation allows the user to interactively select regions of interest and observe the evolving connectivity results during computation. Results are presented from the stochastic generation of over 250,000 fiber steps per iteration at interactive frame rates on consumer-grade graphics hardware.

One of the goals of telemedicine is to enable remote visualization and browsing of medical volumes. There is a need to employ scalable compression schemes and efficient client-server models to obtain interactivity and an enhanced viewing experience. First, we present a scheme that uses JPEG2000 and JPIP (JPEG2000 Interactive Protocol) to transmit data in a multi-resolution and progressive fashion. The server exploits the spatial locality offered by the wavelet transform and packet indexing information to transmit, in so far as possible, compressed volume data relevant to the clients query. Once the client identifies its volume of interest (VOI), the volume is refined progressively within the VOI from an initial lossy to a final lossless representation. Contextual background information can also be made available having quality fading away from the VOI. Second, we present a prioritization that enables the client to progressively visualize scene content from a compressed file. In our specific example, the client is able to make requests to progressively receive data corresponding to any tissue type. The server is now capable of reordering the same compressed data file on the fly to serve data packets prioritized as per the client's request. Lastly, we describe the effect of compression parameters on compression ratio, decoding times and interactivity. We also present suggestions for optimizing JPEG2000 for remote volume visualization and volume browsing applications. The resulting system is ideally suited for client-server applications with the server maintaining the compressed volume data, to be browsed by a client with a low bandwidth constraint.

There is increased use of medical imaging techniques that produce four dimensional (4D) datasets such as fMRI and 3D dynamic echocardiograms. These datasets consume even larger amounts of resources for transmission or storage compared to the traditional 2D data sets. In this paper, we extend the zero tree algorithms, EZW (embedded zero tree coding of wavelet coefficients) and SPIHT (set partitioning in hierarchical trees) to 4D to compress the 4D datasets more efficiently. Integer to integer wavelet transforms scaled by appropriate subband energy weights are used to get lossy to lossless compression. We also investigate the effects of lossy compression on the end result of fMRI analysis.

3D reconstruction of arterial vessels from planar radiographs obtained form multitude of angles around the object has gained increasing interest. The motivating application has been interventional angiography. In these procedures data acquisition, visualization, decision making, and the action are all incorporated in one loop that is time critical. This project encompasses an entire workflow; from system calibration, data acquisition, and image reconstruction, to the effective presentation of the results for the right action.

This paper presents a novel technique for image restoration based on nonlinear interpolative vector quantization (NLIVQ). The algorithm performs nonlinear restoration of diffraction-limited images concurrently with quantization. It is trained on image pairs consisting of an original image and its diffraction-limited counterpart. The discrete cosine transform is used in the codebook design process to control complexity. Simulation results are presented that demonstrate improvements in visual quality and peak signal-to-noise ratio of the restored images.

Two multiframe formulations of Hunt's Poisson maximum a posteriori (MAP) image restoration algorithm (Hunt and Sementilli, 1992; and Hunt 1995) are presented as a means to restore an astronomical object from multiple short-exposure images. This work anticipates accurate wavefront sensor estimates of the optical transfer function of the atmosphere and telescope combination. Theoretical discussion of the algorithms is presented with simulation results. These results indicate that the multiframe algorithms produce significantly better restorations than single frame algorithms. Additionally, our studies demonstrate that the multiframe algorithms presented accomplish super-resolution.

An algorithm based on nonlinear interpolative vector quantization (NLIVQ) is presented which accomplishes image restoration concurrently with image compression. The algorithm is applied to the problem of deblurring noise-free diffraction-limited images by training with a large set of blurred and original image pairs. Simulation results demonstrate a quantitative improvement in images processed by the algorithm, as measured by image peak signal-to-noise ratio (PSNR), as well as a significant improvement in perceived image quality. A theoretical formulation of the algorithm is presented along with a discussion of implementation, training and simulation results.

Multiresolution methods have played a prominent role in image processing since the early eighties. Concurrently, multigrid methods have gained wide acceptance in numerical methods. These two concepts are similar in certain aspects. In this paper, we discuss these similarities in the context of image deconvolution problems. A unified space decomposition formulation linking multiresolution and multigrid methods is presented. In addition, certain earlier deconvolution schemes based on background-detail methods, Grenander's method of sieves etc. are interpreted as variations based on this common formulation.>

A new nonrecurrent associative memory model is proposed. This model is composed of a nonlinear transformation in the spectral domain followed by the association. The Moore-Penrose pseudoinverse is employed to obtain the least squares optimal solution. Computer simulations are done to evaluate the performance of the model. The simulations use one-dimensional speech signals and two-dimensional head/shoulder images. Comparison of the proposed model with the classical optimal linear associative memory and an optimal nonlinear associative memory is presented.