The Medical Imaging Research Laboratory has been in existence since the early eighties, when Dr. Robert Kruger put the pieces in place to set up a medical imaging laboratory that would work closely with the radiology department to develop new, clinically relevant, imaging methods. This presentation will discuss how the department has grown from a small group of 4 faculty housed in the basement of the school of medicine, to a 70 member group housed in a dedicated facility: "The Center for Advanced Medical Technologies". MIRL has become a world renowned center for medical imaging and among its staff are students from all over the world including China, Russia, Poland, Japan, France, India, and Germany. Research areas range from the very theoretical development of 3D reconstruction algorithms to the very clinical development of novel MRA techniques. Current technologies, including MRI, MEG, CT, SPECT and "pseudo" PET comprise the majority of imaging modalities and constitute the basis of research grants from the NIH, the Whitaker Foundation, and the American Heart Association, and research agreements with General Electric, Picker International, Bracco, Nycomed, and StorageTek. Combined, these grants provide an annual budget in excess of $3 million.

The RSH SPECT scanner is a modified conventional SPECT scanner, where a rotating slant-hole (RSH) collimator is used in place of a conventional collimator. The RSH collimator has multiple segments to admit multiplexing of the data collection, and the rotation action reduces the angular sampling requirements of the detector head.

The unusual fully-3D scanning geometry presents challenges for image reconstruction, particularly for attenuation correction. Transmission measurements may be difficult to obtain, so we are investigating approximate approaches that do not use direct knowledge of the attenuation map. We are following the ConTraSPECT concept of Frank Natterer (and developed here by Andy Welch) to obtain an approximate model for the attenuation map using the exponential Radon Transform. Catherine Mennessier will present details of this aspect of the project.

Once the projection data has been converted to a good approximation of exponential Radon transform data, the original emission distribution must be reconstructed. We are investigating direct analytic methods to perform this reconstruction step; we report on some background and progress with this aspect.

Rotating Slant-Hole SPECT uses the principle that the photons emitted from a point in the region of interest will be detected by different regions of the detector depending on the direction of the collimator holes. The advantages of using RSH SPECT are increased sensitivity and improved resolution of the organ scanned. Also, less motion of the detector is sufficient for full tomographic reconstruction of the image as required by OrlovÕs condition. RSH are of two types depending on the number of segments in the collimator (2 segments in the case of BSH: bilateral slant hole collimator and 4 segments in QSH: quadrant slant hole collimator).

It is important to contain the organ to be scanned within the common volume (CV) for accurate reconstruction. The main focus was to develop projection and backprojection routines that will account for all the offsets of the detector. This technique was applied to the images collected using Iowa, Jaszczak and MCAT phantoms for both the collimators (BSH and QSH) and was found to work satisfactorily.

We present a method of accurate attenuation-corrected SPECT reconstruction without transmission measurements for the 2-D parallel gemometry and the 3-D RSH geometry. Our approach differs from previous work such as by Welch et al, as it is based on the new formulation of the exponential Radon transform range conditions, established by Aguilar et al.. A major advantage is that this method extends to fully three-dimensional SPECT imaging such as the RSH SPECT scanner described by Clack et al. (see p. 4). The best parameterized constant attenuation distribution that would account for the observed emission measurements is estimated by maximizing consistency of the normalized data.

We evaluated the method for the 2-D case by performing numerical simulations for the heart and torso. From a non-constant attenuation distribution and the MCAT model of the left ventricle we applied our algorithm to estimate the best ellipse. The resulting simple attenuation distribution was then incorportated into the reconstruction of the activity distribution, and even though it did not closely represent the true attenuation, the mean square error on the reconstruction was less than 3% when compared to the true activity distribution.

Figure: In the top left corner is the emission activity (right and left ventricles) and attenuation map. The attenuation values for the lung, tissue, and bone are 0.039/cm, 0.152/cm, and 0.248/cm, respectively. In the top right corner is the simulated noise-free emission sinogram. The best ellipse-shaped region of constant value equal to 0.161/cm is illustrated in the bottom left hand corner. The estimated emission distribution is shown in the right hand bottom corner.

In recent years dual detector hybrid gamma cameras have become available to perform positron coincidence detection imaging in addition to routine nuclear medicine studies. The primary impetus for this new technology has been oncology imaging using F18-FDG. We have investigated the feasibility of imaging the ultra short-lived myocardial imaging agent rubidium-82 (75 second half-life) using a Picker IRIX system equipped with positron coincidence detection (PCD). The primary issues with this imaging technology are the resolution and sensitivity capabilities. Using a Sr-82/Rb-82 generator provided by Bracco Radiopharmaceuticals, we performed a variety of phantom and animal studies to characterize the imaging performance of this system and radiopharmaceutical. Spatial resolution was compared to Na-22 to identify any resolution degradation from the higher energy positrons of Rb-82. Phantom studies simulating the myocardium both without and with attenuating medium present demonstrated significant image quality, and improvements can be expected with the application of various compensation methods. Based upon the success of the phantom experiments, sequential studies were performed in an animal model to determine optimal acquisition parameters for this short-lived radionuclide when used in vivo. We identified near optimal scanning parameters with an injection of 6 mCi and imaging times as short as 2 minutes, which produced clinically useful results using as few as 4 gantry rotations. Both 2D and 3D acquisition techniques were acquired in these experiments. These results demonstrate that myocardial perfusion imaging with Rb-82 and coincidence detection cameras is a feasible alternative to accepted methods, and with future developments this approach may potentially provide better quality diagnostic images than the current standard.

Figure: Short-axis images of Rb-82 myocardial perfusion imaging performed in a canine. Good image quality was obtained, and the right ventricle can be visualized in some of the slices. The non-uniformities in the myocardial wall of the left ventricle suggest the need for decay correction and compensation for image degrading factors.

To predict the dynamic behavior of the mechanical deformation of the left ventricular myocardium, a finite element method is used to solve the mechanical boundary value problem with known boundary conditions of intra- and extra-ventricular pressures. The cylindrical model is an approximation to the geometry of a thick mid ventricular section of the left ventricule. The finite element analysis code requires that every element and every element corner be assigned a number and the coordinates of every element corner be supplied. Determining and typing in these input data can be difficult and tedious. Mesh generation code automatically generates the mesh fitting the cylindrical geometry, sequences the elements and nodes, defines coordinates, and places all this information in a format that the analysis program can read directly. Generally, the mesh generation code can adjust the density of meshes at different places: the larger the curvature, the denser the meshes in the neighborhood. This will increase the accuracy of the calculations, since those points with extremes of curvatures are crucial to describe certain geometries.

Nowadays, the usual manner to perform attenuation correction in SPECT is to simultaneously acquire transmission data (to reconstruct attenuation map), and emission data (to reconstruct intensity image using attenuation map to correct for attenuation effect). Recently investigators have studied the potential of using the consistency between emission data and the corresponding transmission data to perform attenuation correction. This technique allows one to compute the attenuation distribution without acquiring transmission data. However, because the system is highly under-determined, it is necessary to apply a priori information to the attenuator. This study presents three different ways to define the attenuator. The first method uses an affine deformation of an initial boundary. The second is based on the assumption that the unknown object is defined by a spline. The third technique utilizes the same information as the second except that the untruncated part of the transmission sinogram is used. Results of a heart simulation study are presented. It is shown that the first two methods provide equivalent results because these methods give an uniform estimation of a non-uniform attenuator. These results show that the third method provides better results since information from the real transmission sinogram is included.

Figure: (a) original object; (b) reconstruction using affine transform; (c) reconstruction using spline curve; (d) reconstruction using spline curve and untruncated part of transmission sinogram.

For the inversion of SPECT data, given by the attenuated Radon Transform

one usually needs some a priori information about the attenuation map m. If both f and m are unknown, then some algorithm is required to invert the nonlinear equation. The main idea of the talk is to approximate the the operator R by a bilinear operator R,

with linear operator A and bilinear operator B and to apply iterative methods for linear problems to the bilinear equation.

I would like to report on the main results and conclusions of the P31 MRS study. The goal is to study the peripheral contribution to fatigue in Multiple Sclerosis patients. The study raw data was obtained using an H decoupled P31 surface coil from arm muscle during a maximum voluntary exercise. P31 Data was obtained in vivo from a single voxel and localized by the MRI signal from H in the muscle. The FID was sampled and processed both manually in the frequency domain and in the time domain using an iterative pre-filtered non-linear least-squares (Marquardt-Levenberg variable projection), which is initialized by a linear prediction SVD and perturbed by a simulated annealing procedure to search for the global minima.

Muscle metabolite concentration, rate of recovery, pH, etc. information is obtained and monitored as time marches on, and contrasted for the three MS patient groups and the normal group. SAS statistics is applied for the grouped data to look for any statistical significance amongst the different groups.

The noisy and inherently weak P31 signal and coil sensitivity, lack of an absolute internal metabolite reference, and other grouping factors are some of the challenges in this study.

Centrally ordered phase-encoding (PE) has been used in 2D MRI and MRA to reduce artifacts by destroying the periodicity of cyclic motion. Unlike 2D imaging, there are many possible methods of ordering the central views in 3D imaging. In a set of simulations, different types of PE order applied to 3D gradient echo and fast-spin echo is studied. An elliptical centric PE order is shown to be more resistant to motion artifacts than a number of other centric PE techniques. Proper implementation of the centric elliptical order requires consideration not only of the sampling matrix size but also of the relative field of views (FOVs) in the two PE directions because this order is based on k-space radii. An additional advantage of the centric elliptical PE order is that the most important views are acquired first, allowing for zero replacement of the last views that can be used for limiting of the acquisition time or partial k-space updating.

The goal of coil design is simple: maximize the signal intensity and minimize the noise. In order to better understand the effect of coil design on signal-to-noise ratio (SNR), an experiment was designed to qualitatively analyze the effects of coil losses on SNR. The experiment consists of comparing the SNR in coils constructed of different materials and sizes. Coils were constructed of 10 AWG wire, copper braid, etched, and double-sided etched. The double-sided coils were etched on two different substrates-- one substrate with a dielectric constant of 2.5 and the other substrate with a dielectric of 9.6. Square surface coils were constructed by each method to be 12.7 cm x 12.7 cm and 7.6 cm x 7.6 cm. Phantom studies with a spin echo sequence will be used to acquire images and SNR will be compared for each coil. This presentation will present the results of the experiment and what has been learned.

Hyperthermia is a cancer therapy used in conjunction with radiation therapy, which has been shown to have a higher success rate in treating cancer than does radiation alone. The goal of hyperthermia treatments is to raise the tumor tissue temperature to therapeutic levels, while limiting thermal energy in healthy tissue to acceptable levels. Accurate temperature measurement is critical to patient treatment. Previously, tissue temperature has been read discretely using thermocouples that are invasively inserted into the patient. It has been shown that temperature can be read using magnetic resonance imaging techniques, specifically the proton chemical shift which is detected using phase difference imaging. Magnetic field drift is a concern when using phase difference imaging and is currently being investigated. An agar phantom is being used to provide a known temperature standard that can be used to calibrate and verify non-invasive temperature measurements throughout the region of interest. Developing a technique to account for drift will enable full temperature fields to be measured non-invasively. This will allow the practical, clinical use of a scanned focused ultrasound system that has been developed for use in the MR scanner.

The organization of white matter fiber bundles in the brain can be mapped out using diffusion-tensor MRI methods. Maps of the white matter organization may significantly improve both the understanding of neurophysiology and the planning of neurosurgical procedures. An estimate of the diffusion-tensor is obtained by combining image measurements with diffusion-weighting in six different directions. Since the diffusion of fibrous tissues such as white matter have a strong directional component, the locations of the white matter fibers can be mapped out by mapping the diffusion anisotropy as a function of position (see Figure). The orientation of the fiber bundles is estimated by using the first principle component of the diffusion tensor. Currently, diffusion-tensor images are obtained using a diffusion-weighted echo-planar imaging method. The design of the diffusion-tensor MRI pulse sequence, the tensor image calculation and white matter visualization techniques will be described. Potential applications of this technique will be outlined.

T2 weighted image Anisotropy image showing

areas of white matter

We present reconstruction results from helical cone-beam CT data, obtained using a simple and fast algorithm, which we call the CB-SSRB algorithm. This algorithm combines the single-single rebinning method of PET imaging with the weighting schemes of spiral CT algorithms. The reconstruction is approximate but can be performed using 2-D multi-slice fan-beam filtered backprojection. The quality of the results is surprisingly good and far exceeds what one might expect, even when the pitch of the helix is large. In particular, with this algorithm comparable quality is obtained using helical cone-beam data with a normalized pitch of 10 to standard spiral CT reconstruction with a normalized pitch of 2.

Figure: Rebinning description. (a) Each fan-beam ray is assigned the value of a cone-beam (CB) ray emitted by a source directly above or below the virtual fan source. (b) and (c) The selected CB ray passes through the mid-point M of the intersection of the fan-line with the region-of-interest (ROI).

Whole body SPECT with a helical scanning orbit has a potential impact in cancer detection. In order to correct for attenuation and avoid trunction or projection data, an off-center cone-beam imaging geometry is investigated for simultaneous transmission and emission data aquisition. A data sufficiency condition is used to develop a requirement for the pitch of the helical orbit so that the projection data are complete. A single helical orbit is shown not to be able to provide sufficient projection data; at least two detectors are required to acquire a complete data set using a helical orbit. If a two-detector, half cone-beam imaging system is used and the two detectors are positioned face-to-face, the requirement for complete data measurement is that the pitch of the helical orbit is 2a, where a is the detector size projected on the axis of rotation. Computer simulations are performed to verify the orbit pitch requiremant. The iterative ML-EM algorithm is used to reconstruct the image. The proposed two-detector, half cone-beam, helical SPECT system is shown to be practical, and artifact-free reconstructions are obtained.

An iterative Bayesian reconstruction algorithm based on the total variation (TV) norm constraint is proposed. The motivation for using TV regularization is that it is extremely effective for recovering edges of images. The TV norm minimization, introduced in 1992, was shown to be effective for restoring blurred images with a Gaussian noise model and was demonstrated to be effective for noise suppression and edge preservation. The images were diffused according to a set of nonlinear anisotropic diffusion partial differential equations, which suffered from computational difficulties. This paper extends the TV norm minimization constraint to the field of SPECT image reconstruction with a Poisson noise model. The regularization norm is included in the ML-EM (maximum likelihood expectation maximization) algorithm. The partial differential equation approach is not utilized here. Reconstructions of computer simulations and patient data show that the proposed algorithm has the capacity to smooth the noise and maintain sharp

edges without introducing over/under shoots and ripples around the edges.

A variety of effects may make the apical region appear colder than it actually is when performing cardiac SPECT image reconstructions. In this paper, simulations were performed using fan-beam collimators to study the effect of geometric point response correction and attenuation correction on the apical region in cardiac SPECT. The images were reconstructed using the iterative ordered-subset expectation maximization (OS-EM) algorithm, in which warping-and-rotation based projector/backprojector pairs were employed. The smoothing introduced by interpolations in the algorithm was also studied. It was shown that reconstructions without performing the geometric point response correction did not further decrease the activity of the apical region. Use of the attenuation map reconstructed from the truncated transmission data when performing the attenuation correction might result in an artificially cool apical region. Comparison of the images reconstructed with the OS-EM algorithm and the filtered back-projection algorithm demonstrated that the partial volume effect related to the interpolations in the OS-EM algorithm did not introduce visible activity decrease in the apical region for the cardiac orientation used.

A new data acquisition protocol for dynamic SPECT imaging was used. The detector was rotated slowly around the object performing only 3 rotations during the whole dynamic study. The resulting projections of such acquisition were not consistent in time. The dynamic sequence was reconstructed then where one dynamic image corresponded to only one projection. Such a highly underestimated reconstruction process-each image was reconstructed only from one projection-was possible by using of a factor model. The Euclidean norm between the model and the data was used as an objective function. Due to usage of a factor model this function was not linear, and standard minimization methods, like conjugant gradient, failed to give satisfactory results. Simulated annealing was used to minimize the objective function. Either the factors or the factor coefficients were modified in each iteration and a stochastic decision of which to modify, factors or factor coefficients, was made based on the average absolute value of the partial derivative around the current configuration. Each minimization required about 200 iterations for 64x64 pixel image with 183 projections and 64 bins per projection. The method gave good results for computer simulations of the MCAT phantom with 3 and 4 factors.

A 4D ordered-subsets maximum a posteriori (OSMAP) algorithm for dynamic SPECT is described. It uses a temporal prior that constrains each voxel's behavior in time to obey a compartmental model. No a priori limitations on kinetic parameters are applied; rather, the parameter estimates evolve as the algorithm iterates to a solution. The estimated parameters are also used to model changes in the activity distribution as the camera rotates, avoiding artifacts due to data inconsistencies between angles. The algorithm was evaluated for dynamic cardiac SPECT imaging with teboroxime using a two compartment model. Canine study results demonstrated qualitative improvements for OSMAP compared to OSEM. In a simulation experiment, 100 noise realizations were reconstructed using both OSEM and OSMAP. Population means and standard deviations of regional kinetic parameters were compared: True values: k21=0.8, k12=0.4 min-1; OSEM: k21=0.77 +/-.06, k12=.41 +/-.06; OSMAP: k21=.76 +/-.04, k12=.41 +/-.05. OSMAP provided parameter estimates with significantly lower standard deviations than OSEM at similar bias.

Figure: Transaxial reconstructed images of a dynamic teboroxime SPECT canine study for 4 iterations of OSEM (top row) and OSMAP (bottom row). Time frames, from left to right, were 1.5 min, 5 min, 10 min, and 15 min. The OSMAP reconstruction prior was limited to an 800 voxel region containing the heart and blood pool for this slice. The OSMAP reconstruction has markedly reduced noise effects as compared to the OSEM reconstruction.

Measurement of dynamic anatomical and physiological processes with PET, MRI, or SPECT provides a wealth of clinically relevant information for a wide array of disease states. In many of the methods, knowledge of the blood input function is essential in the computation of the physiologically relevant kinetic parameters. This work is concerned with developing methods to accurately estimate these kinetic parameters blindly; that is, without use of a directly measured blood input function. Instead, only measurements of the output functions - the tissue time-activity curves - are used. The approach is to factor the input function out of the equations used to fit the dynamic data by employing blind estimation methods similar to ones developed in the field of telecommunications. Such an approach may reduce or eliminate the need for an input function in cardiac, brain, and oncological dynamic scans. Since directly measuring the blood input function is relatively invasive in many cases, there is considerable motivation to develop methods to estimate kinetic parameters without requiring measurements of the blood input.

Here, a deterministic blind estimation approach is applied to simulated data representative of dynamic PET cerebral blood flow imaging with 15O water, dynamic SPECT cardiac perfusion imaging with 99mTc-teboroxime, and dynamic MRI cardiac imaging using bolus injections of paramagnetic contrast.

Figure: Simulated SPECT teboroxime curves. The blood input (curve with highest maximum) and four noisy time-activity curves are shown. Also shown are curves that were created from kinetic parameters found with the blind estimation method (lines marked with o's).

The inability for current magnetic resonance angiography techniques to visualize small cerebral vessels frequently results in x-ray catheter angiography techniques being used clinically. However, magnetic resonance angiography offers several advantages over catheter angiography, including three dimensional depiction of the vasculature, lower cost, and lower morbidity. Efforts to improve small vessel visualization have been based on both physical techniques (e.g., magnetization transfer, paramagnetic contrast agents, increased image resolution) and post-processing techniques (zero-filled interpolation, and vessel enhancement algorithms). Vessel enhancement algorithms have been developed previously for enhancing visualization of vessels using non-lin- ear line detection methods. These methods assumed that vessels were pixel sized lines. However, with increased acquisition and reconstruction resolution, many important vessels are larger than a pixel in diameter and thus have decreased visualization when these algorithms are applied. To overcome this problem we have implemented multi-scale vessel enhancement algorithms using gaussian subsampling of the images. Vessel changes are examined in terms of signal-difference-to-noise ratios. The changes induced by the enhancement algorithms are compared to changes induced by both application of magnetization transfer and injection of a paramagnetic contrast agent.

One approach used in the imaging of vascular anatomy in MRI is the so called time-of-flight (TOF) method whereby the background tissue within the excitation volume is saturated and the signal is derived from "fresh" spins flowing into the volume. A common TOF technique employed clinically is the multiple overlapping thin slab acquisition (MOTSA) which offers a compromise between the strong in-flow enhancement of a 2D slice acquisition and the superior signal to noise of 3D slab imaging. However, the in-slab saturation of the flow inherent in the 3D approach, the imperfect slab selection profile, and patient motion during acquisition all contribute to a distinct artifact associated with MOTSA, the slab boundary artifact (SBA). We have been working with and extending the Sliding Interleaved ky (SLINKY) technique which seeks to remove the SBA by modifying the normal k-space sampling order while maintaining the sampling efficiency and resolution of MOTSA. SLINKY can produce image volumes free of the SBA of traditional MOTSA, yet is itself susceptible to a type of SBA which arises from phase errors due to the changes in k-space sampling order. We will discuss our efforts to determine the optimum k-space sampling order, our introduction of oversampling to overcome the technique's remaining SBA, and the extension of the method to projection reconstruction MRA.

Multiple overlapping thin slab acquisition (MOTSA) techniques, in general, provide a mechanism to trade-off the noise reduction properties of 3D MR acquisition with the blood signal enhancement of thin slice 2D MR acquisition. Unfortunately, MOTSA techniques show a significant tendency to produce slab boundary related intensity variations. The SLinding Interleaved Ky (SLINKY) acquisition technique developed by Liu and Rutt has been shown to reduce this artifact, at the expense of causing ghosting within the slices. Recently, Roberts has demonstrated that these artifacts can be reduced by highly overlapping measurements taken in the center of k-space. In this presentation we show that projection acquisition techniques can also be applied to this problem. We have demonstrated that projection reconstruction techniques can be applied to multiple slab acquisition. When used in conjunction with sliding interleaved acquisition, the resulting images appear to completely eliminate the slab boundary artifact. Unfortunately, SLIPR and projection reconstruction techniques in general, are very sensitive to a knowledge of the center of rotation which in turn depends directly on the local, DC, magnetic field. If the primary field is not homogeneous, regions of the image can be significantly blurred.

One formidable challenge when diagnosing cardiac disease is ascertaining whether perfusion defects represent viable myocardial tissue or infarcted tissue. It is known that if wall motion and wall thickening can be determined, then there is a high probability that myocardial viability can be determined. Recently, it has been shown that magnetic resonance imaging (MRI) can precisely and accurately measure local three-dimensional myocardial deformation. However, at present it is a challenge for single photon emission computed tomography (SPECT) to achieve similar results when measuring three-dimensional myocardial deformation because of poor statistics and poor resolution. Present techniques of edge tracking are effective for dynamically following wall thickness and wall motion in a short axis slice. However, they are inadequate for following rotation out of the plane. Here, particular attention is given to the development of a mechanical model of the beating heart, which is based on an elastic boundary value problem. The model provides additional information in the estimation of myocardial deformation using gated SPECT and gated MRI data. This model assumes that the heart is an incompressible hyperelastic media containing fiber bundle sheaths of a specified helical orientation. Attenuation corrected gated SPECT data was fitted to the mechanical model and the results were compared with fits of gated MRI data. The results compared well with strain calculations made directly from SPAMM MRI data. In clinical practice this technique should provide better representations of the dynamic mechanical deformation of the myocardium when analyzing gated SPECT data than the technique of following dynamically detected edge points.