Image Representation Framework Research Articles

Improved low-dose positron emission tomography image reconstruction using deep learned prior

Positron emission tomography (PET) is a promising medical imaging technology that provides non-invasive and quantitative measurement of biochemical process in the human bodies. PET image reconstruction is challenging due to the ill-poseness of the inverse problem. With lower statistics caused by the limited detected photons, low-dose PET imaging leads to noisy reconstructed images with much quality degradation. Recently, deep neural networks (DNN) have been widely used in computer vision tasks and attracted growing interests in medical imaging. In this paper, we proposed a maximum a posteriori (MAP) reconstruction algorithm incorporating a convolutional neural network (CNN) representation in the formation of the prior. Rather than using the CNN in post-processing, we embedded the neural network in the reconstruction framework for image representation. Using the simulated data, we first quantitatively evaluated our proposed method in terms of the noise-bias tradeoff, and compared with the filtered maximum likelihood (ML), the conventional MAP, and the CNN post-processing methods. In addition to the simulation experiments, the proposed method was further quantitatively validated on the acquired patient brain and body data with the tradeoff between noise and contrast. The results demonstrated that the proposed CNN-MAP method improved noise-bias tradeoff compared with the filtered ML, the conventional MAP, and the CNN post-processing methods in the simulation study. For the patient study, the CNN-MAP method achieved better noise-contrast tradeoff over the other three methods. The quantitative enhancements indicate the potential value of the proposed CNN-MAP method in low-dose PET imaging.

Iterative PET Image Reconstruction Using Convolutional Neural Network Representation.

PET image reconstruction is challenging due to the ill-poseness of the inverse problem and limited number of detected photons. Recently, the deep neural networks have been widely and successfully used in computer vision tasks and attracted growing interests in medical imaging. In this paper, we trained a deep residual convolutional neural network to improve PET image quality by using the existing inter-patient information. An innovative feature of the proposed method is that we embed the neural network in the iterative reconstruction framework for image representation, rather than using it as a post-processing tool. We formulate the objective function as a constrained optimization problem and solve it using the alternating direction method of multipliers algorithm. Both simulation data and hybrid real data are used to evaluate the proposed method. Quantification results show that our proposed iterative neural network method can outperform the neural network denoising and conventional penalized maximum likelihood methods.

Deep p-Fibonacci scattering networks

Recently, the use of neural networks for image classification has become widely spread. Thanks to the availability of increased computational power, better performing architectures have been designed, such as the Deep Neural networks. In this work, we propose a novel image representation framework exploiting the Deep p-Fibonacci scattering network. The architecture is based on the structured p-Fibonacci scattering over graph data. This approach allows to provide good accuracy in classification while reducing the computational complexity. Experimental results demonstrate that the performance of the proposed method is comparable to state-of-the-art unsupervised methods while being computationally more efficient.

A Novel Image Representation Framework Based on Gaussian Model and Evolutionary Optimization

We propose a novel image representation framework based on Gaussian model and evolutionary optimization (EO). In this framework, image patches are categorized into smooth and nonsmooth ones, and the two categories are treated distinctively. For a smooth patch, we formulate it as the summation of a direct component and a variation component (VC). We observe that the values of all VCs in an image can be well fitted by a Gaussian distribution, according to which we present an efficient reconstruction approach based on maximizing the logarithm a posteriori probability. For a nonsmooth patch, we introduce the mechanism of EO to solve a combinatorial optimization over a principal component analysis dictionary. In addition, we develop two approaches for estimating the coefficients of the atoms. Experiment results demonstrate that the proposed framework obtains the state-of-the-art results in several image inverse problems.

Discriminant Kernel Assignment for Image Coding.

This paper proposes discriminant kernel assignment (DKA) in the bag-of-features framework for image representation. DKA slightly modifies existing kernel assignment to learn width-variant Gaussian kernel functions to perform discriminant local feature assignment. When directly applying gradient-descent method to solve DKA, the optimization may contain multiple time-consuming reassignment implementations in iterations. Accordingly, we introduce a more practical way to locally linearize the DKA objective and the difficult task is cast as a sequence of easier ones. Since DKA only focuses on the feature assignment part, it seamlessly collaborates with other discriminative learning approaches, e.g., discriminant dictionary learning or multiple kernel learning, for even better performances. Experimental evaluations on multiple benchmark datasets verify that DKA outperforms other image assignment approaches and exhibits significant efficiency in feature coding.

A Tuned Mesh-Generation Strategy for Image Representation Based on Data-Dependent Triangulation

A mesh-generation framework for image representation based on data-dependent triangulation is proposed. The proposed framework is a modified version of the frameworks of Rippa and Garland and Heckbert that facilitates the development of more effective mesh-generation methods. As the proposed framework has several free parameters, the effects of different choices of these parameters on mesh quality are studied, leading to the recommendation of a particular set of choices for these parameters. A mesh-generation method is then introduced that employs the proposed framework with these best parameter choices. This method is demonstrated to produce meshes of higher quality (both in terms of squared error and subjectively) than those generated by several competing approaches, at a relatively modest computational and memory cost.

A highly-effective incremental/decremental Delaunay mesh-generation strategy for image representation

A flexible mesh-generation framework for image representation based on Delaunay triangulations is proposed. By fixing the various degrees of freedom available within this framework, two mesh-generation methods, known as ID1 and ID2, are derived. These two methods are shown to perform extremely well, producing meshes of significantly higher quality than state-of-the-art schemes at relatively low computational cost. Furthermore, the ID1 and ID2 methods each provide a mechanism whereby mesh quality can be increased (or decreased) in return for a corresponding increase (or decrease) in computational cost. Lastly, we demonstrate that one component of our proposed methods, called bad-point replacement, can be used as a postprocessing optimization step that, when added to other previously-proposed mesh-generation methods, yields meshes of much greater quality.

Multilinear Supervised Neighborhood Embedding of a Local Descriptor Tensor for Scene/Object Recognition

In this paper, we propose to represent an image as a local descriptor tensor and use a multilinear supervised neighborhood embedding (MSNE) for discriminant feature extraction, which is able to be used for subject or scene recognition. The contributions of this paper include: 1) a novel feature extraction approach denoted as the histogram of orientation weighted with a normalized gradient (NHOG) for local region representation, which is robust to large illumination variation in an image; 2) an image representation framework denoted as the local descriptor tensor, which can effectively combine a moderate amount of local features together for image representation and be more efficient than the popular existing bag-of-feature model; and 3) an MSNE analysis algorithm, which can directly deal with the local descriptor tensor for extracting discriminant and compact features and, at the same time, preserve neighborhood structure in tensor-feature space for subject/scene recognition. We demonstrate the performance advantages of our proposed approach over existing techniques on different types of benchmark database such as a scene data set (i.e., OT8), face data sets (i.e., YALE and PIE), and view-based object data sets (COIL-100 and ETH-80).

Logarithmic Adaptive Neighborhood Image Processing (LANIP): Introduction, Connections to Human Brightness Perception, and Application Issues

A new framework for image representation, processing, and analysis is introduced and exposed through practical applications. The proposed approach is called logarithmic adaptive neighborhood image processing (LANIP) since it is based on the logarithmic image processing (LIP) and on the general adaptive neighborhood image processing (GANIP) approaches, that allow several intensity and spatial properties of the human brightness perception to be mathematically modeled and operationalized, and computationally implemented. The LANIP approach is mathematically, computationally, and practically relevant and is particularly connected to several human visual laws and characteristics such as: intensity range inversion, saturation characteristic, Webers and Fechners laws, psychophysical contrast, spatial adaptivity, multiscale adaptivity, morphological symmetry property. The LANIP approach is finally exposed in several areas: image multiscale decomposition, image restoration, image segmentation, and image enhancement, through biomedical materials and visual imaging applications.

Tomographic reconstruction using an adaptive tetrahedral mesh defined by a point cloud

Medical images in nuclear medicine are commonly represented in three dimensions as a stack of two-dimensional images that are reconstructed from tomographic projections. Although natural and straightforward, this may not be an optimal visual representation for performing various diagnostic tasks. A method for three-dimensional (3-D) tomographic reconstruction is developed using a point cloud image representation. A point cloud is a set of points (nodes) in space, where each node of the point cloud is characterized by its position and intensity. The density of the nodes determines the local resolution allowing for the modeling of different parts of the image with different resolution. The reconstructed volume, which in general could be of any resolution, size, shape, and topology, is represented by a set of nonoverlapping tetrahedra defined by the nodes. The intensity at any point within the volume is defined by linearly interpolating inside a tetrahedron from the values at the four nodes that define the tetrahedron. This approach creates a continuous piecewise linear intensity over the reconstruction domain. The reconstruction provides a distinct multiresolution representation, which is designed to accurately and efficiently represent the 3-D image. The method is applicable to the acquisition of any tomographic geometry, such as parallel-, fan-, and cone-beam; and the reconstruction procedure can also model the physics of the image detection process. An efficient method for evaluating the system projection matrix is presented. The system matrix is used in an iterative algorithm to reconstruct both the intensity and location of the distribution of points in the point cloud. Examples of the reconstruction of projection data generated by computer simulations and projection data experimentally acquired using a Jaszczak cardiac torso phantom are presented. This work creates a framework for voxel-less multiresolution representation of images in nuclear medicine.