Tool For Image Reconstruction Research Articles

Siamese Cooperative Learning for Unsupervised Image Reconstruction From Incomplete Measurements.

Image reconstruction from incomplete measurements is one basic task in imaging. While supervised deep learning has emerged as a powerful tool for image reconstruction in recent years, its applicability is limited by its prerequisite on a large number of latent images for model training. To extend the application of deep learning to the imaging tasks where acquisition of latent images is challenging, this article proposes an unsupervised deep learning method that trains a deep model for image reconstruction with the access limited to measurement data. We develop a Siamese network whose twin sub-networks perform reconstruction cooperatively on a pair of complementary spaces: the null space of the measurement matrix and the range space of its pseudo inverse. The Siamese network is trained by a self-supervised loss with three terms: a data consistency loss over available measurements in the range space, a data consistency loss between intermediate results in the null space, and a mutual consistency loss on the predictions of the twin sub-networks in the full space. The proposed method is applied to four imaging tasks from different applications, and extensive experiments have shown its advantages over existing unsupervised solutions.

Multifrequency Black Hole Imaging for the Next-generation Event Horizon Telescope

The Event Horizon Telescope (EHT) has produced images of the plasma flow around the supermassive black holes in Sgr A* and M87* with a resolution comparable to the projected size of their event horizons. Observations with the next-generation Event Horizon Telescope (ngEHT) will have significantly improved Fourier plane coverage and will be conducted at multiple frequency bands (86, 230, and 345 GHz), each with a wide bandwidth. At these frequencies, both Sgr A* and M87* transition from optically thin to optically thick. Resolved spectral index maps in the near-horizon and jet-launching regions of these supermassive black hole sources can constrain properties of the emitting plasma that are degenerate in single-frequency images. In addition, combining information from data obtained at multiple frequencies is a powerful tool for interferometric image reconstruction, since gaps in spatial scales in single-frequency observations can be filled in with information from other frequencies. Here we present a new method of simultaneously reconstructing interferometric images at multiple frequencies along with their spectral index maps. The method is based on existing regularized maximum likelihood (RML) methods commonly used for EHT imaging and is implemented in the eht-imaging Python software library. We show results of this method on simulated ngEHT data sets as well as on real data from the Very Long Baseline Array and Atacama Large Millimeter/submillimeter Array. These examples demonstrate that simultaneous RML multifrequency image reconstruction produces higher-quality and more scientifically useful results than is possible from combining independent image reconstructions at each frequency.

Offsets between X-Ray and Radio Components in X-Ray Jets: The AtlasX

The X-ray emission mechanism of powerful extragalactic jets—which has important implications for their environmental impacts—is poorly understood. The X-ray/radio positional offsets in the individual features of jets provide important clues. Extending previous work in Reddy et al., we present a detailed comparison between X-ray maps, deconvolved using the Low-count Image Reconstruction and Analysis tool, and radio maps of 164 components from 77 Chandra-detected X-ray jets. We detect 94 offsets (57%), with 58 new detections. In FR II–type jet knots, the X-rays peak and decay before the radio in about half the cases, disagreeing with the predictions of one-zone models. While a similar number of knots lack statistically significant offsets, we argue that projection and distance effects result in offsets below the detection level. Similar deprojected offsets imply that X-rays could be more compact than radio for most knots, and we qualitatively reproduce this finding with a “moving-knot” model. The bulk Lorentz factor (Γ) derived for knots under this model is consistent with previous radio-based estimates, suggesting that kiloparsec-scale jets are only mildly relativistic. An analysis of the X-ray/radio flux ratio distributions does not support the commonly invoked mechanism of X-ray production from inverse Compton scattering of the cosmic microwave background, but does show a marginally significant trend of declining flux ratio as a function of the distance from the core. Our results imply the need for multi-zone models to explain the X-ray emission from powerful jets. We provide an interactive list of our X-ray jet sample at http://astro.umbc.edu/Atlas-X.

Determination of patient-specific airway stent fit using novel 3D reconstruction measurement techniques: a 4-year follow-up of a patient.

Managing complex benign airway disease is a major challenge in interventional pulmonology. With the introduction of additive manufacturing in the medical field, patient-specific (PS) implants are an innovate prospect for airway management. Historically, stents were oversized to resist migration. However, the optimal degree and impact of stent oversizing remains unclear. The ability to design stents based on computed tomography (CT) invites opportunity to understand sizing. Here, we report a novel three-dimensional (3D) image reconstruction tool to quantify fit repeatedly over time. Analysis of CT imaging before and after successive stent implants in a single patient with different areas of stenosis and malacia was done. Nine PS airway stents over 4 years (five left mainstem and four right mainstem) were studied. The distance between the airway model and stent was calculated. The CT images were correlated to stent designs in CloudCompare software (v2.10-alpha) for novel analysis. Heat map was exported depicting the distances between the airway and the stent to the clinician's prescribed stent model. Corresponding histograms containing distances, mean, and standard deviation were reported. It is possible to measure stent fit based on heat map quantification on patient imaging. Observation of the airway over time and stent change suggests that the airway became more open over time requiring increased stent diameters. The ability to design and measure stent fit over time can help quantify the utility and impact of PS silicone airway stent. The airway appears to display plasticity such that there is notable change in stent prescription over time.

Applications of ultrasound image reconstruction using deep learning

Deep learning has gained tremendous popularity as a tool for ultrasound beamforming and image reconstruction. In previous work, we trained deep neural networks (DNNs) to estimate the echogenicity of a medium, to improve acoustical and electronic signal-to-noise ratio (SNR) in channel data, and to detect targeted microbubbles nondestructively for real-time ultrasound molecular imaging. Here, we present several advancements to each application. First, we compare the speckle- and noise-reducing performance of DNNs trained with simple linear Field II simulations of photographic images versus that of DNNs trained with full wave finite-difference time domain numerical simulations containing realistic abdominal walls and the resulting image degradation artifacts. We further extend our nondestructive molecular imaging DNN to incorporate spatiotemporal information using an extended simulation study to increase specificity for stationary bound microbubbles and further improve nondestructive molecular imaging.

Electrophysiological confrontation of Lead-DBS-based electrode localizations in patients with Parkinson’s disease undergoing deep brain stimulation

Microelectrode recordings (MERs) are often used during deep brain stimulation (DBS) surgeries to confirm the position of electrodes in patients with advanced Parkinson’s disease.The present study focused on 32 patients who had undergone DBS surgery for advanced Parkinson’s disease. The first objective was to confront the anatomical locations of intraoperative individual MERs as determined electrophysiologically with those determined postoperatively by image reconstructions. The second aim was to search for differences in cell characteristics among the three subthalamic nucleus (STN) subdivisions and between the STN and other identified subcortical structures.Using the DISTAL atlas implemented in the Lead-DBS image reconstruction toolbox, each MER location was determined postoperatively and attributed to specific anatomical structures (sensorimotor, associative or limbic STN; substantia nigra [SN], thalamus, nucleus reticularis polaris, zona incerta [ZI]). The STN dorsal borders determined intraoperatively from electrophysiology were then compared with the STN dorsal borders determined by the reconstructed images. Parameters of spike clusters (firing rates, amplitudes – with minimum amplitude of 60 μV -, spike durations, amplitude spectral density of β-oscillations) were compared between structures (ANOVAs on ranks).Two hundred and thirty one MERs were analyzed (144 in 34 STNs, 7 in 4 thalami, 5 in 4 ZIs, 34 in 10 SNs, 41 others). The average difference in depth of the electrophysiological dorsal STN entry in comparison with the STN entry obtained with Lead-DBS was found to be of 0.1 mm (standard deviation: 0.8 mm). All 12 analyzed MERs recorded above the electrophysiologically-determined STN entry were confirmed to be in the thalamus or zona incerta. All MERs electrophysiologically attributed to the SN were confirmed to belong to this nucleus. However, 6/34 MERs that were electrophysiologically attributed to the ventral STN were postoperatively reattributed to the SN. Furthermore, 44 MERs of 3 trajectories, which were intraoperatively attributed to the STN, were postoperatively reattributed to the pallidum or thalamus.MER parameters seemed to differ across the STN, with higher spike amplitudes (H = 10.64, p < 0.01) and less prevalent β-oscillations (H = 9.81, p < 0.01) in the limbic STN than in the sensorimotor and associative subdivisions. Some cells, especially in the SN, showed longer spikes with lower firing rates, in agreement with described characteristics of dopamine cells. However, these probabilistic electrophysiological signatures might become clinically less relevant with the development of image reconstruction tools, which deserve to be applied intraoperatively.

A Pixel-Dependent Finite Element Model for Spatial Frequency Domain Imaging Using NIRFAST

Spatial frequency domain imaging (SFDI) utilizes the projection of spatially modulated light patterns upon biological tissues to obtain optical property maps for absorption and reduced scattering. Conventionally, both forward modeling and optical property recovery are performed using pixel-independent models, calculated via analytical solutions or Monte-Carlo-based look-up tables, both assuming a homogenous medium. The resulting recovered maps are limited for samples of high heterogeneity, where the homogenous assumption is not valid. NIRFAST, a FEM-based image modeling and reconstruction tool, simulates complex heterogeneous tissue optical interactions for single and multiwavelength systems. Based on the diffusion equation, NIRFAST has been adapted to perform pixel-dependent forward modeling for SFDI. Validation is performed within the spatially resolved domain, along with homogenous structured illumination simulations, with a recovery error of &lt;2%. Heterogeneity is introduced through cylindrical anomalies, varying size, depth and optical property values, with recovery errors of &lt;10%, as observed across a variety of simulations. This work demonstrates the importance of pixel-dependent light interaction modeling for SFDI and its role in quantitative accuracy. Here, a full raw image SFDI modeling tool is presented for heterogeneous samples, providing a mechanism towards a pixel-dependent SFDI image modeling and parameter recovery system.

Neural networks as an applied tool for ultrasound beamforming and image reconstruction

In recent years, the computer science community has made available its artificial intelligence tools to the public at-large. Freely available online courses, software libraries, and databases have made it easy to incorporate artificial intelligence, and specifically machine learning, into many research areas. In medical imaging, this has led to a proliferation of machine learning-based research methodologies in image reconstruction tasks, whether appropriate or not. In this lecture, we review machine learning as a tool for ultrasound beamforming. We discuss the selection of an appropriate problem to be solved with machine learning, survey common and popular implementations of neural networks, and describe best practices and common pitfalls. We visit a few misconceptions about machine learning in ultrasound imaging and examine the recent literature in the field of ultrasound beamforming and image reconstruction.

Applications, promises, and pitfalls of deep learning for fluorescence image reconstruction.

Deep learning is becoming an increasingly important tool for image reconstruction in fluorescence microscopy. We review state-of-the-art applications such as image restoration and super-resolution imaging, and discuss how the latest deep learning research could be applied to other image reconstruction tasks. Despite its successes, deep learning also poses substantial challenges and has limits. We discuss key questions, including how to obtain training data, whether discovery of unknown structures is possible, and the danger of inferring unsubstantiated image details.

GPU-Based Iterative Medical CT Image Reconstructions

The algebraic reconstruction technique (ART) is an iterative algorithm for CT (i.e., computed tomography) image reconstruction that delivers better image quality with less radiation dosage than the industry-standard filtered back projection (FBP). However, the high computational cost of ART requires researchers to turn to high-performance computing to accelerate the algorithm. Alas, existing approaches for ART suffer from inefficient design of compressed data structures and computational kernels on GPUs. Thus, this paper presents our CUDA-based CT image reconstruction tool based on the algebraic reconstruction technique (ART) or cuART. It delivers a compression and parallelization solution for ART-based image reconstruction on GPUs. We address the under-performing, but popular, GPU libraries, e.g., cuSPARSE, BRC, and CSR5, on the ART algorithm and propose a symmetry-based CSR format (SCSR) to further compress the CSR data structure and optimize data access for both SpMV and SpMV_T via a column-indices permutation. We also propose sorting-based global-level and sorting-free view-level blocking techniques to optimize the kernel computation by leveraging different sparsity patterns of the system matrix. The end result is that cuART can reduce the memory footprint significantly and enable practical CT datasets to fit into a single GPU. The experimental results on a NVIDIA Tesla K80 GPU illustrate that our approach can achieve up to 6.8x, 7.2x, and 5.4x speedups over counterparts that use cuSPARSE, BRC, and CSR5, respectively.

Neural Network Image Reconstruction for Magnetic Particle Imaging

We investigate neural network image reconstruction for magnetic particle imaging. The network performance depends strongly on the convolution effects of the spectrum input data. The larger convolution effect appearing at a relatively smaller nanoparticle size obstructs the network training. The trained single-layer network reveals the weighting matrix consisted of a basis vector in the form of Chebyshev polynomials of the second kind. The weighting matrix corresponds to an inverse system matrix, where an incoherency of basis vectors due to a low convolution effects as well as a nonlinear activation function plays a crucial role in retrieving the matrix elements. Test images are well reconstructed through trained networks having an inverse kernel matrix. We also confirm that a multi-layer network with one hidden layer improves the performance. The architecture of a neural network overcoming the low incoherence of the inverse kernel through the classification property will become a better tool for image reconstruction.

Multi-layer strategy and reconstruction model with low rank and local rank regularizations for single image super-resolution

Low rank method has been proven to be a powerful tool for image reconstruction. However, because of being hard to find sufficient similar patches for those high gradient patches which include rich texture details and strong edge information, the classical low rank method always destroys the critical image details and fails to preserve the edge structures. To tackle the above mentioned problems, in this paper, we propose a new multi-layer strategy and reconstruction model which combines the low rank and local rank regularization for image super resolution. In the proposed method, firstly, considering that each layer of the decomposed image shares the similar edge or texture information, which contributes to searching enough and more accurate similar patches to faithfully characterize each exemplar patch, we can decompose the input image into multi-layer images to resolve the problem of destroying critical image details. Next, we incorporate the local rank regularization into the low rank model to develop a new reconstruction model to effectively reconstruct the image details and sharpen the edge structures simultaneously. The experimental results on various images validate the superiority of the proposed method in terms of visual quality and objective indices.

TOOLBOX FOR 3D MODELLING AND IMAGE RECONSTRUCTION IN ELECTRICAL CAPACITANCE TOMOGRAPHY

Electrical Capacitance Tomography is used to visualize a spatial distribution of electric permittivity in a tomographic sensor. ECT is able to create even thousands of frames per second which is suitable for application in the industry, e.g. monitoring of multiphase flows or material mixing. A tool for sensor modelling and image reconstruction is needed in order to develop improved solutions and to better understand phenomena in ECT. A software for 2D and 2D modelling is developed in the Division of Nuclear and Medical Electronics. In this paper a Matlab toolbox called ECTsim for 3D modelling is presented.

PET and SPECT of Neurobiological Systems

1805 Objectives A PET image reconstruction tool (Q.Clear®) has been recently introduced that allows more iterations of the raw PET data than the standard OSEM image reconstruction, to enable contrast recovery without amplifying noise. Here we qualitatively and semi-quantitatively compared Q.Clear® with the standard OSEM image reconstruction available on our clinical PET/CT scanners (GE Discovery 600 and 690) using patient data. Methods PET datasets from 30 previously acquired clinical research scans (16 68Ga-DOTA TATE, 11 18F-FPPRGD2, 1 18F-FLT, 2 18F-FSPG) were reconstructed using Q.Clear® (25 iterations with 3 different regularization strength or “beta” values) and compared to standard reconstruction for Discovery 600 (2 iterations and 32 subsets; with PSF modeling) and 690 (2 iterations and 24 subsets; with PSF modeling) scanners. Each of the 4 reconstructions/scan were reviewed blindly by 2 experienced nuclear medicine physicians and scored 1-5 (1 poor, 5 excellent quality). The SUVmax of the smallest lesion detected in each patient was also measured for comparison. The paired Welch t-test was used for statistical analysis. Results Significant differences (P Conclusions Q.Clear® with a higher number of iterations maintains image quality. Visual analysis confirmed that overall image quality is improved and may allow for improved detection of small lesions as compared to the image quality from the standard OSEM algorithm. Research Support GE Healthcare

Model Based Filtered Backprojection Algorithm: A Tutorial.

People have been wandering for a long time whether a filtered backprojection (FBP) algorithm is able to incorporate measurement noise in image reconstruction. The purpose of this tutorial is to develop such an FBP algorithm that is able to minimize an objective function with an embedded noise model. An objective function is first set up to model measurement noise and to enforce some constraints so that the resultant image has some pre-specified properties. An iterative algorithm is used to minimize the objective function, and then the result of the iterative algorithm is converted into the Fourier domain, which in turn leads to an FBP algorithm. The model based FBP algorithm is almost the same as the conventional FBP algorithm, except for the filtering step. The model based FBP algorithm has been applied to low-dose x-ray CT, nuclear medicine, and real-time MRI applications. Compared with the conventional FBP algorithm, the model based FBP algorithm is more effective in reducing noise. Even though an iterative algorithm can achieve the same noise-reducing performance, the model based FBP algorithm is much more computationally efficient. The model based FBP algorithm is an efficient and effective image reconstruction tool. In many applications, it can replace the state-of-the-art iterative algorithms, which usually have a heavy computational cost. The model based FBP algorithm is linear and it has advantages over a nonlinear iterative algorithm in parametric image reconstruction and noise analysis.

Singular value decomposition for the truncated Hilbert transform

Starting from a breakthrough result by Gelfand and Graev, inversion of the Hilbert transform became a very important tool for image reconstruction in tomography. In particular, their result is useful when the tomographic data are truncated and one deals with an interior problem. As was established recently, the interior problem admits a stable and unique solution when some a priori information about the object being scanned is available. The most common approach to solving the interior problem is based on converting it to the Hilbert transform and performing analytic continuation. Depending on what type of tomographic data are available, one gets different Hilbert inversion problems. In this paper, we consider two such problems and establish singular value decomposition for the operators involved. We also propose algorithms for performing analytic continuation.

Design, implementation, and optimization of MEGAlib's image reconstruction tool Mimrec

MEGAlib, the Medium-EnergyGamma-ray Astronomylibrary, is a toolset to simulate and analyze data from gammaray detectors. An integral part of MEGAlib is its imaging tool Mimrec, which performs list-mode-likelihood image deconvolution. Mimrec has to handle data from coded masks, Compton cameras, and pair conversion telescopes with different response representations, on different imaging grids, with different deconvolution algorithms, etc. This versatility requires a highly modular and object-oriented design to avoid overhead and code redundancy. In addition, since some applications require close to real-time image reconstruction, great care has to be taken to optimize the library.

Cerebrospinal Fluid Flow Dynamics in the Central Nervous System

Cine-phase-contrast-MRI was used to measure the three-dimensional cerebrospinal fluid (CSF) flow field inside the central nervous system (CNS) of a healthy subject. Image reconstruction and grid generation tools were then used to develop a three-dimensional fluid-structure interaction model of the CSF flow inside the CNS. The CSF spaces were discretized using the finite-element method and the constitutive equations for fluid and solid motion solved in ADINA-FSI 8.6. Model predictions of CSF velocity magnitude and stroke volume were found to be in excellent agreement with the experimental data. CSF pressure gradients and amplitudes were computed in all regions of the CNS. The computed pressure gradients and amplitudes closely match values obtained clinically. The highest pressure amplitude of 77Pa was predicted to occur in the lateral ventricles. The pressure gradient between the lateral ventricles and the lumbar region of the spinal canal did not exceed 132Pa (~1mmHg) at any time during the cardiac cycle. The pressure wave speed in the spinal canal was predicted and found to agree closely with values previously reported in the literature. Finally, the forward and backward motion of the CSF in the ventricles was visualized, revealing the complex mixing patterns in the CSF spaces. The mathematical model presented in this article is a prerequisite for developing a mechanistic understanding of the relationships among vasculature pulsations, CSF flow, and CSF pressure waves in the CNS.

Cerebrospinal Fluid Flow in the Normal and Hydrocephalic Human Brain

Advances in magnetic resonance (MR) imaging techniques enable the accurate measurements of cerebrospinal fluid (CSF) flow in the human brain. In addition, image reconstruction tools facilitate the collection of patient-specific brain geometry data such as the exact dimensions of the ventricular and subarachnoidal spaces (SAS) as well as the computer-aided reconstruction of the CSF-filled spaces. The solution of the conservation of CSF mass and momentum balances over a finite computational mesh obtained from the MR images predict the patients' CSF flow and pressure field. Advanced image reconstruction tools used in conjunction with first principles of fluid mechanics allow an accurate verification of the CSF flow patters for individual patients. This paper presents a detailed analysis of pulsatile CSF flow and pressure dynamics in a normal and hydrocephalic patient. Experimental CSF flow measurements and computational results of flow and pressure fields in the ventricular system, the SAS and brain parenchyma are presented. The pulsating CSF motion is explored in normal and pathological conditions of communicating hydrocephalus. This paper predicts small transmantle pressure differences between lateral ventricles and SASs (approximately 10 Pa). The transmantle pressure between ventricles and SAS remains small even in the hydrocephalic patient (approximately 30 Pa), but the ICP pulsatility increases by a factor of four. The computational fluid dynamics (CFD) results of the predicted CSF flow velocities are in good agreement with Cine MRI measurements. Differences between the predicted and observed CSF flow velocities in the prepontine area point towards complex brain-CSF interactions. The paper presents the complete computational model to predict the pulsatile CSF flow in the cranial cavity.

Full 3-D cluster-based iterative image reconstruction tool for a small animal PET camera

Iterative reconstruction methods are commonly used to obtain images with high resolution and good signal-to-noise ratio in nuclear imaging. The aim of this work was to develop a scalable, fast, cluster based, fully 3-D iterative image reconstruction package for our small animal PET camera, the miniPET. The reconstruction package is developed to determine the 3-D radioactivity distribution from list mode type of data sets and it can also simulate noise-free projections of digital phantoms. We separated the system matrix generation and the fully 3-D iterative reconstruction process. As the detector geometry is fixed for a given camera, the system matrix describing this geometry is calculated only once and used for every image reconstruction, making the process much faster. The Poisson and the random noise sensitivity of the ML-EM iterative algorithm were studied for our small animal PET system with the help of the simulation and reconstruction tool. The reconstruction tool has also been tested with data collected by the miniPET from a line and a cylinder shaped phantom and also a rat.