Measurement of the average power spectra of perfusion lung scans shows that the distribution of macroaggregated albumin in the walls of the lung lobules behaves as a fractal object. Measures related to the power spectral slope for all views in a study, mean slope and slope standard deviation, have a high correlation with the visual interpretation of the radiologist. Application of appropriate fractal models reveals that the surface dimension of normal lung parenchyma range between 2.7 and 3.0.

Mammographic parenchymal pattern has been associated with risk of developing breast cancer. Currently, these patterns are classified subjectively by radiologists according to "Wolfe grade", as N, P1, P2, or DY. Such broad classifications have limited applicability for the assessment of subtle changes which may occur, for example, in long term studies of women at high risk for breast cancer. For such work, a consistent, quantitative, observer-independent method of characterization is required. We have been developing such a method based on the use of "fractals". We have applied techniques of calculating fractal dimension to digitized mammograms. To evaluate our technique, we are measuring the degree of correlation between the fractal classification of images and their mammographic patterns as assessed by radiologists. Preliminary results have shown significant correlation between the fractal-based and radiologist-based assessments.

An important goal in medical imaging is to increase the accuracy of visual detection of small abnormal regions. The presence of scatter in the image degrades spatial resolution by introducing long tails to the point-spread function. We show in this paper that linear deconvolution can be used to improve the performance of the human observer in the two-hypothesis detection task. Also, we investigate the effect that linear grey-scale mappings have on the human observer performance. We demonstrate that they help the human observer in the detection task and can be used sequentially with deconvolution to yield a better performance.

Parametric model-based approaches to 3-D reconstruction of vessels overcome the inherent problem of underdeterminancy in reconstruction from limited views by incorporating a priori knowledge about the structure of vessels and about the measurement statistics. In this paper, we describe two extensions to the parametric approach. First, we consider the problem of reconstruction from a pair of bi-plane angiograms that are acquired at different projection angles. Since bi-plane angiography systems are widely available, this is a practical measurement geometry. The patient may move between acquisitions, so we have extended our model to allow for object translation between the first and second pair of projections. Second, we describe how to accurately estimate the dimensions of a aneurysm from the dual-biplane angiogram. We applied the new algorithm to four synthetic angiograms (projection angles 0°, 20°, 90°, and 110°) of a vessel with a small aneurysm and an eccentric stenosis. The angiograms were corrupted by additive noise and background structure. Except near the top and bottom of the aneurysm, the estimated cross sections of the aneurysm and stenosis agree very well with the true cross sections.

Early experience of new forms of adaptive filtering for ultrasound speckle reduction and parametric imaging, using off-line conventional digital processing, has been sufficiently encouraging to warrant attempts to build fast on-line implementations for purposes of clinical trials. This paper describes the hardware under development. The feasibility of implementing specific algorithms in real-time (video rate) at low cost was demonstrated using a hybrid digital/analogue system in which the video signal (from any scanner) is sampled to 256 points per line and passed sequentially through a series of shift registers, in order to derive a 5 by 5 window of values which surrounds the image point to be processed. These 25 video signals are then used as inputs to an analogue processor, which provides the filtered output. For greater accuracy, flexibility, spatial resolution and freedom from image artefacts, a slower system has been constructed around a commercially available digital image processor. This system has been chosen to act as the main clinical test device. Whole images may be processed in a few tens of seconds and specified regions, if small enough, may be processed at rates approaching real-time.

Motion artifacts in magnetic resonance images result from displacement of "isochromats" during the imaging interval. This motion is categorized into two classes: motion within a Free Induction Decay (FID) and motion between FID excitations. Furthermore, local regions in the spatial data being measured generally undergo unique motion profiles for each class, and each class may not consist of the same local segmentation. Models for each class are presented using a general digital signal processing (DSP) format and conversions from real world parameters to this normalized DSP domain are given. A multilevel motion model is then given, suggesting a layered approach to removing the spatial and class variant distortion functions. Finally, a method and simulation is presented for removing one of the more troublesome of these artifact layers, global patient movement.

The paper describes ongoing work in the isolation, visualization, and qualification of pelvic anatomies imaged through Magnetic Resonance Imaging.

Reconstructing three-dimensional objects, both surface and interior intensities, from serial cross-sections not only makes slicing from any arbitrary angle possible but also provides accurate quantitative information for various purposes. However, the only technique available for intensity interpolation is the straightforward linear interpolation method which is technically unsound and groundless. In this paper, an intensity interpolation method for two regions of interest lying on two consecutive cross-sections is proposed. Although the intensity is interpolated linearly, the corresponding points involved in interpolation are carefully selected. The result is compared with the straightforward linear interpolation method.

We describe a method for matching computed tomography (CT) and/or magnetic resonance (MR) medical image data with digital subtraction angiograms (DSA). The tomographic data set is processed and rendered to yield a volumetric image which is displayed stereoscopically from the same viewpoint at which the DSA images were acquired. The vessels are merged on top of the 3D rendered volumes and the composite images are viewed stereoscopically on a PC-based workstation. The workstation has facilities allowing the user to manipulate a. cursor or vector in stereo within a defined 3D coordinate system and to make accurate 3D measurements within the imaged volume. While surgical planning can be conducted from analysis of the separate tomographic images and individual bi-plane angiograms, such an approach is tedious since as many as 400 separate images from separate modalities may be acquired for a single patient. Furthermore, it is sometimes difficult to fully appreciate the anatomy of the surrounding structures from individual CT and MR slices of the 3D volume. All image processing and rendering is performed on a PIXAR Image Computer, a SIMD (single instruction, multiple data) channel array processor with a 48 - bit deep image buffer. PIXAR's Chap Volumes rendering library is employed in the rendering process. The stereoscopic workstation consists of a high performance color monitor which displays the left and right eye views of the data sequentially at 120 Hz via a liquid crystal polarizing shutter. The operator views the image via left and right eye circularly polarized glasses.

In this paper, we propose a new method to construct the joint probability model from a specified first-order distribution and correlation structure. The construction procedure can be interpreted in two ways: (1) It embodies the maximum entropy principle, or (2) It is considered as a correlated non-Gaussian source generated by a nonlinear transformation from a correlated Gaussian source. Its stochastic properties [mean,correlation] and information-theoretic properties [entropy, rate-distortion bound] are examined. An example for the lognormal distribution is given to illustrate the construction process and the characteristics of the source. This approach should remove the limitations imposed by earlier methods, and make for more realistic modeling of medical images.

This paper describes an algorithm for the restoration of echocardiographic sequences of several consecutive heart beats. It is based on the estimation of the parameters of a quasi-periodic signal model. This model is fully characterized by a one-period-long reference signal and a warping function that defines the mapping between the reference and observed time-scales. Given an initial reference template, an optimal warping function is determined using dynamic programming. This function is optimal in the sense that it minimizes the mean square error between the warped template and the measured noisy signal. A reference signal estimate is then formed by averaging several cardiac cycles with reference to a normalized time-scale. The efficiency of this procedure, which may be iterated, is demonstrated quantitatively using quasi-periodic noisy test data. This method is then applied to the improvement of M-mode echocardiograms.

The morphology of the coronary arteries in single-view angiograms is studied on the basis of densitometric and geometric continuity properties along the vessel segment. The vessel contour is automatically identified by use of a sequential tracking algorithm. The tracking algorithm is based on the assumption of geometric continuation and similarity between the current and the next incremental section along the vessel. A mathematical description of the vessel contour that comprises the lumen width, direction, and intensity as functions of distance along the vessel centerline is generated. The accuracy of the contour representation is quantitatively verified with a synthetic vessel image. The algorithm is applied to the analysis of human coronary cineangiograms and digital subtraction angiograms. Fourier analysis is performed to determine the spatial frequency characteristics of the vessel contour functions. It is shown that, for a vessel of average 9 pixels in lumen width, 90% signal energy of the contour functions is within the spatial frequency range of 0 - 0.09 pixel-1. The result of this study should be useful in analyzing coronary arterial abnormalities and tortuosity. The methodology should also be applicable to vessel contour representation and vessel image compression.

A novel filter was designed that preserves signal edges better than a linear or a median filter when matched to attain the same amount of noise smoothing. This filter was modeled after well-known properties of neurons in the visual cortex that are only sensitive to visual stimuli of lines tilted in specific directions. Using a similar directional sensitivity, the presence of signal edges in the filter window can be detected and the smoothing applied only in directions that do not cross edges. In the presence of white, Gaussian noise, the minimum SNR must be at least 1/2 to attain superiority in terms of edge preservation of this new directional filter, for lower SNR values its performance may deteriorate ungracefully and a linear filter is the clear choice.

This work discusses the development of an image segmentation algorithm that utilizes fractional Brownian motion as a fractal texture model. Additionally a second feature, the correlation coefficient of the regression analysis associated with the fractal dimension computation, is also used to assess which regions of the radiograph are well described by the fractional Brownian motion model and hence, are more likely to be bone. The resulting three dimension feature space consisting of the local fractal dimension, radiograph intensities and regression analysis correlation coefficient was partitioned to identify each pixel of the radiographs as being either bone, teeth, or a boundary between these entities. The results obtained in this study indicate the potential utility of a textural based algorithm for segmentation of dental radiographs.

This paper describes our new high-speed image reconstruction display. This display reconstructs high-quality images using a hybrid algorithm which combines DCT and BTC methods, and communicates with the host computer at high speed through a BMC. To test the display's performance, we developed a prototype using this display and a main frame computer (Fujitsu M-340). Compressed images stored in the host's database are transferred to the displays via the BMC, and reconstructed using its local function. Experiment showed the display to be effective: It takes less than 5 seconds for a 10 to 1 compressed image to be transferred and reconstructed the image, while taking 10 seconds for the original image.

This paper presents work in progress concerning an automatic system for the 3D reconstruction and representation of the cerebral vessels. It is based on a separate delineation of the blood vessels in two stereo images. First, we extract blood vessel segments from the image and subsequently we use those high level primitives to guide the stereoscopic matching process. Therefore, we make extensive use of domain specific knowledge like the orientation, thickness and intensity of blood vessels.

A true color imaging technique has been developed for analysis of microscopic fluorochromic bone biopsy images to quantify new bone growth. The technique searches for specified colors in a medical image for quantification of areas of interest. Based on a user supplied training set, a multispectral classification of pixel values is performed and used for segmenting the image. Good results were obtained when compared to manual tracings of new bone growth performed by an orthopedic surgeon. At a 95% confidence level, the hypothesis that there is no difference between the two methods can be accepted. Work is in progress to test bone biopsies with different colored stains and further optimize the analysis process using three-dimensional spectral ordering techniques.

We report on work in progress on segmentation and presentation of three-dimensional vascular morphology on 3D MR and Doppler Ultrasound angiograms. The segmentation strategy is twofold. First, we enhance blood-vessel-like structures using 3D non-linear filters. Second, these images can be directly displayed or used as input to medium-level processes. The medium-level filters are cast in the framework of energy minimization using deformable models.

In this study the automatic detection of clusters of microcalcifications in digital mammograms was investigated. A local area thresholding technique was employed to segment all potential microcalcifications from the normal breast structure. These objects were then analysed using size, shape and gradient measures to extract clusters of microcalcifications. The results for a set of 25 mammographic regions, each 5.76 X 5.76 cm2 in area, show that the computer system can achieve 100% true positive cluster detection with a false positive rate of 12%.

In this paper, an expert vision system is proposed which integrates knowledge from diverse sources for tomographic image segmentation. The system miinicks the reasoning process of an expert to divide a tomographic brain image into semantically meaningful entities. These entities can then be related to the fundamental biomedical processes, both in health and in disease, that are of interest or of importance to health care research. The images under study include those acquired from x-ray CT (Computed Tomography), MRI (Magnetic Resonance Imaging), and PET (Positron Emission Tomography). Given a set of three (correlated) images acquired from these three different modalities at the same slicing level and angle of a human brain, the proposed system performs image segmentation based on (1) knowledge about the characteristics of the three different sensors, (2) knowledge about the anatomic structures of human brains, (3) knowledge about brain diseases, and (4) knowledge about image processing and analysis tools. Since the problem domain is characterized by incomplete and uncertain information, the blackboard architecture which is an opportunistic reasoning model is adopted as the framework of the proposed system.

Multiresolution data structures, such as image pyramids, have many different uses in image processing and analysis. We present a segmentation algorithm which uses image pyramids for efficient segmentation of 3D images. The segmented objects are converted from a voxel to a polyhedral surface representation. Results of experiments on synthetic and actual CT data are presented, and future enhancements to the algorithm are discussed.

One of the major problems in 3-D volume reconstruction from magnetic resonance imaging (MRI) is the difficulty in automating the classification of soft tissues. Because of the complicated soft tissue structures revealed by MRI, it is not easy to segment the images with simple algorithms. MRI can obtain multiple images from the same anatomical section with different pulse sequences, with each image having different response characteristics for each soft tissue. Using the gray level distributions of soft tissues, we have developed two statistical classifiers that utilize the image context information based on the Markov Random Field (MRF) image model. One of the classifiers classifies each voxel to a specific tissue type and the other estimates the partial volume of each tissue within each voxel. Since the voxel sizes of tomographic images are finite and the measurements from tissue boundaries represent the mixture of multiple tissue types, it is preferable that the classifier should not classify each voxel in all-or-none fashion; rather, it should be able to tell the percentage volume of each class in each voxel for the better visualization of the prepared 3-D dataset. The paper presents the theoretical basis of the algorithms and experimental evaluation results of the classifiers in terms of classification accuracy, as compared to the conventional maximum likelihood classifier.

To overcome the typical difficulties for the extraction of biomedical structures in microscopical images, a knowledge-based system has been designed with the aim of performing a model-driven segmentation towards automatic structure recognition. The work deals with the characterization of cell populations in human embryonal and foetal organs, in particular blood vessels and nervous cells. By using a-priori known description (in linguistic form) about the structures to be detected, a thresholding segmentation is initially used for locating darker pixels (markers). Then gradient extraction, region-growing, or other techniques are invoked around the area bounding markers. The computation of some specific attributes of the analyzed areas drives the recognition process allowing to maintain regions where structures are detected, and discard regions where structure presence is not verified. The segmentation and recognition process is controlled by a production system whose rules are activated on the basis of the input data, the progressive results, and the information (provided by the user) about the structures to be localized. Some results are presented to the user, obtained by changing processing parameters. They correspond to recognized maps with different reliability factors.

Motion artifacts caused by moving anatomy degrade images acquired by means of digital subtraction angiography. It is in principle possible to remove these artifacts through image registration by means of geometrical transformations of properly calibrated images, but there are known difficulties with this approach. In this paper we consider two of these difficulties, the presence of x-ray scatter and, more significantly, the inherent absence of sufficient information to identify correct transformations. We perform experiments using a novel phantom to illustrate these difficulties. The phantom is a balloon that is inflated within the lumen of a coronary artery during percutaneous transluminal angioplasty. We employ two registration techniques, a previously described technique to compensate for gross, simple motion, and a novel one to compensate for smaller, more intricate motion.

In this article we will present work in progress concerning a knowledge-based system for the labeling of the coronary arteries on single projections. The approach is based on a gradual refinement of the interpretation results, starting from the detection of blood vessel center lines, the extraction of bar-like primitives and the connection into blood vessel segments. In this paper we will focus on the final stage which is the labeling of the delineated blood vessel segments. In contrast with most existing approaches which are mainly based on a sequential labeling of the vessels starting from the most important segment, our system uses a constraint satisfaction technique. Mainly, because most anatomical knowledge can be easily formalized as constraints on local attributes such as position, greyvalue, thickness and orientation and as constraints on relations between blood vessel segments such as "left of" or "in same direction". Anatomical models are developed for the Left Coronary Artery in standard RAO and LAO views. In general, only 1-2 interpretations are left, which is an encouraging result if you take into account that for some projections there is a considerable overlap between vessel segments.

A previously reported semi-analytical expression for calculating the point spread function (PSF) of singly scattered x-ray radiation is shown to exhibit excellent correspondence with Monte Carlo-determined PSFs where only single scatter was allowed to occur. A neural network was trained to transform the semi-analytic single scatter PSF to the PSF for both single and multiple scatter. Using comparisons between the semi-analytic/neural network generated and the Monte Carlo generated PSFs, excellent agreement was found. The technique proposed here appears to be an accurate method for rapidly calculating the scatter PSF from three physical parameters, and such a formulism should be useful in algorithms designed to correct for scattered radiation effects in digital images.

A method is described for preprocessing projection data prior to image reconstruction in single-photon emission computed tomography. The projection data of the desired spatial distribution of emission activity is blurred by the point-response function of the collimator that is used to define the range of directions of gamma-ray photons reaching the detector. The point-response function of the collimator is not stationary, but depends on the distance from the collimator to the point. Conventional methods for deblurring collimator projection data are based on approximating the actual depth-dependent point-response function by a spatially-invariant blurring function, so that deconvolution methods can be applied independently to the data at each angle of view. The method described in this paper is based on Fourier analysis of the multi-angular data set as a whole, using special depth-dependent characteristics of the Fourier coefficients to achieve spatially-variant inverse filtering of the data in all views simultaneously. Preliminary results are presented for simulated data with a simple collimator model.

On a PACS network, it is desirable to use a non-destructive image compression technique in order both to minimize the storage and to improve the transmission speed. However, if the effect of noise in images is not taken into account, the expected degree of compression may not be achieved. We have studied some radiological images with different levels of noise using various decomposition methods incorporated with Huffman and Lempel-Ziv coding. When more correlations exist between pixels, these techniques can be made more efficiently. On the other hand, the additional noise disrupts the correlation between adjacent pixels and leads to a less compressed result. Hence, prior to a systematic compression in a PAC system two main issues need to be addressed: a) the true information range which exists in a specific type of radiological image and b) the costs and benefits of compression for the PACS.

Automated image matching has important applications, not only in the fields of machine vision and general pattern recognition, but also in modern diagnostic and therapeutic medical imaging. Image matching, including the recognition of objects within images as well as the combination of images that represent the same object or process using different descriptive parameters, is particularly important when complementary physiological and anatomical images, obtained with different imaging modalities, are to be combined. Correlation analysis offers a powerful technique for the computation of translational, rotational and scaling differences between the image data sets, and for the detection of objects or patterns within an image. Current correlation-based approaches do not efficiently deal with the coupling of the registration variables, and thus yield iterative and computationally-expensive algorithms. A new approach is presented which improves on previous solutions. In this new approach, the registration variables are de-coupled, resulting in a much less computationally expensive algorithm. The performance of the new technique is demonstrated in the matching of MRI and PET scans, and in an application of pattern recognition in linear accelerator images.

Quantitative interpretation of functional images (PET or SPECT) is hampered by poor spatial resolution, low counting statistics and, for many tracers, low contrast between different brain structures of interest. Further, normal tracer distributions can be severely distorted by such gross pathologies as stroke, tumor and dementia. Hence, the complementary anatomical information provided by CT or MRI is essential for accurate and reproducible regional analysis of functional data. We have developed methods for the three-dimensional integration and simultaneous display of image volumes from MRI and PET. PET data was collected from an older Therascan 3-slice scanner with 12 mm resolution and a 15-slice Scanditronix PC-2048 system having 5-6 mm resolution in each dimension. MRI data was obtained from a Philips 1.5 Tesla Gyroscan scanner. The image volumes were loaded into a PIXAR 3-D image computer for simultaneous display. A general algorithm for finding the optimal transformation between two ensembles of equivalent points was implemented and investigated through simulation studies. Using a locally-developed 3-D image/graphics analysis package, equivalent points in the two image volumes were identified, either manually or via an adjustable computerized volume-of-interest (VOI) atlas. The MRI data were then re-sampled along planes parallel to the PET planes and the two volumes overlaid using opacity-weighted composition. Arbitrary oblique planes through the two volumes were obtained in interactive sessions.

The present paper proposes a method for interpolation of grayvalued images such as electron microphotographs for medical use. The method consists of four parts. The first is the extraction of feature regions from a pair of grayvalued key frame images. The second is making correspondence to determine disparities between each feature regions. The third is the interpolation of these disparities from one key frame to the other. The last is the generation of inbetweening images. In this part texture mapping is used. This method can be applied to not only medical uses but also computer animation and visual communications. This method can also interpolate color images without any additional techniques.

Intra- and inter-observer variability in the perception of radiographic findings by radiologists and the resulting uncertainty in the presence or absence of these findings complicates the process of verifying the accuracy of expert systems that use this data. To validate these expert systems, techniques must be developed to obtain a "true" set of findings for each case in the database. Several techniques are presented for obtaining these data sets. These methods may have utility in other domains where the input to an expert system consists of features which have significant variability in perception.

Adaptive Histogram Equalization (AHE) has been applied to high resolution digital chest radiographs to provide contrast enhancement. The method provides good contrast in uniform areas of the image, e.g. the lung field, but in so doing both overenhances noise and produces an artifact at boundaries between high density and low density regions. The artifact, which appears as a band of very low contrast data spanning such a boundary, has the effect of suppressing structural information. Although it is known that the problem of overenhancing noise is controlled by the algorithm known as Contrast Limited Adaptive Histogram Equalization (CLAHE), the boundary artifact is not removed by this technique. This paper concentrates on the boundary artifact. We present a method for processing a chest radiograph by means of background subtraction prior to applying the CLAHE algorithm which reduces contrast at high/low density boundaries and thus permits contrast enhancement free of both noise and boundary artifacts.

Many techniques have been proposed to enhance radiographic images and each of them may be "optimal" depending upon the circumstances. However, the problem confronting the radiologist or the physician is which enhancement to use and how to select the parameters when he wants a specified feature to be emphasized. At the University of Ottawa, our research work is oriented towards automatic context-dependent enhancements. Our approach attempts to match the three phases involved in viewing a radiograph: getting a global impression, analysing the objects and the local features and focussing on the image perturbations. In this paper, we report on enhancements to support the two first phases in the case of chest radiographs as well as on the applicability of grey level reversal transformations.

This paper describes a fully automated system for quantitatively analyzing digital coronary angiograms and ventriculograms. Angiographic and ventriculographic image sequences are initially acquired in a purely digital format and stored on a video-rate digital disk system. Left ventriculograms and coronary angiograms are analyzed by this system which functionally is divided into three distinct categories: 1) Image Sequence Review, 2) Coronary Vasculature Analysis, and 3) Left Ventricular Analysis. The software system described integrates the quantitative analysis functions required in a cardiac catheterization laboratory into a single, menu-driven package. The angiograms and ventriculograms displayed, enhanced, and analyzed are in a digital format. All analyses are performed automatically with provisions for user intervention and editing. The algorithms and models upon which this package is based are described.

The relationships of three objective performance measures for the evaluation of image reconstruction algorithms are presented. Two of them are physical metrics - distance and relative error. The third is a numerical observer performance measure. In a demonstration project, three reconstruction algorithms are evaluated on the basis of a specific task using the above measures. Based on a similarity measure between the rank orderings of 30 reconstructions by the performance measures, relative error is seen to have a significant similarity to numerical observer performance. The similarity between distance and numerical observer performance is considerably less. The following hypothesis appears to be worthy of further investigation: judicious use of simple physical metrics, in place of more elaborate and complicated measures, can in certain circumstances be an attractive alternative method for evaluating image reconstruction algorithms.

One of the initial steps in the analysis of 3-D/4-D images is Segmentation, which entails partitioning the images into relevant subsets such as object and background. In this paper, we present a multidimensional segmentation algorithm to extract object surfaces from Computed Tomography (CT) and Magnetic Resonance Imaging (MRI) scans. The algorithm is formulated in the framework of Mathematical Morphology. We propose the Generalized Morphological operators for segmentation in multidimensions. Apriori knowledge of the approximate location of the object surface is commu-nicated to the algorithm via the definition of the Search Space. The algorithm uses this definition of the Search Space to obtain the Surface Candidate elements. The search space specification reduces the computational cost and increases the reliability of the detected features. The detected surface is represented as a hierarchical combination of patches. Initial results obtained by using the algorithm to segment the Brain from cranial MRI scans is presented.

Digital image analysis techniques and statistical discriminant analysis are used to identify mitotic cells in the biopsies of cervical dysplasia. This dysplasia is a preinvasive stage of cervical cancer. Parameters extracted from each cell include areal and geometrical measures, optical density statistics, bending energy, and various statistical texture features. The discriminant analysis attempts to classify mitotic cells from those in interphase on the basis of the image analysis output data vectors using pre-scored classifications. The first comprehensive trial of the imaging technique for a dataset of 32 cervical cells showed that 88% of the cells could be properly identified automatically. A fast, accurate and robust expert imaging system is ultimately desired for the classification of cervical cancer and other types of malignancy.

The scatter point spread function (PSF) is determined for various geometries in diagnostic two dimensional projection imaging, using Monte Carlo (MC) techniques designed to take into account polychromatic spectra (at 100 kVp) and multiple scattering directions and histories. With the knowledge of the primary photon fraction, a total normalized system PSF (primary plus scatter) is derived for each case using analytical and numerical techniques. Numerical Hankel transformation of the PSF profile provides a frequency domain filter that is inverted and applied to experimentally acquired images of a homogeneous lucite phantom matching the MC simulation geometry and technique. Frequency domain processing of the scatter degraded images, followed by inverse transformation, results in images with the scatter component accurately removed in most cases, except for a DC offset. A semi-analytic neural network derived PSF is also used on experimental images, demonstrating similar results as the MC derived filter with the added benefit of time efficient implementation on a case by case basis.

For the past few years we have been conducting physical,psychophysical, and clinical evaluations of the Toshiba computed radiography system (TCR) [1-4]. The research described in this paper grew out of the seemingly conflicting information derived from the physical and psychophysical evaluations conducted on the TCR by our group.

An early diagnosis of osteoporosis is needed to develop an adequate strategy against this disease. The loss of bone in osteoporosis is, among other clinical signs, radiographically visible as thinning of the trabecular bone pattern. The trabeculae and the radiolucent areas in between are ill defined. Computer aided image processing and pattern recognition techniques can assist in diagnosing subtle alterations in the trabecular bone pattern. In previous studies the feasibility of a computer aided procedure for the quantification and description of the trabecular pattern in osteoporotic patients based on a set of 7 parameters was already shown. The study described in this paper gives an evaluation of some characteristics of these parameters used so far to describe the features of the trabecular pattern. The correlation between parameters found for images showing the trabecular pattern was compared with results for a series of randomly selected images. It was concluded that the parameters used to describe the radiographic trabecular bone pattern show a systematic correlation which is different from the other class of images. More research is still to be done before the underlying system in the relation between the parameters can be appreciated to its full extent.

We have investigated whether or not the temporal correlation in time series of angiograms can be utilized to improve the efficiency of reversible data compression. Reversible methods for motion-compensated and non motion-compensated temporal decorrelation in combination with spatial decorrelation have been considered. The conclusion is that the interframe decorrelation techniques considered so far are not substantially better than intraframe techniques.

Twelve radiologists each reviewed a randomly selected mixture of 50 pairs of conventional radiographic images (CRI) and contrast-enhanced digital radiographic images (CEDRI) of neonates with either necrotizing enterocolitis (NEC) or no suspected intra-abdominal abnormality (normal controls). Each image was rated as to overall suspicion of NEC. The data for the NEC and control groups were analyzed and compared to a multiple-review "gold standard" as determined by an experienced pediatric radiologist. For both the CRI and CEDRI, the radiologists tended to underrate the suspicion of NEC in those neonates with NEC and overrate the suspicion of NEC in the control neonates without NEC. Digital manipulation of the original CRI resulted in an increased sensitivity but decreased specificity for the radiographic diagnosis of NEC.

Morphological operators such as erosion, dilation, opening and closing have been used for image processing. These classical operators modify an image feature based only on the size and polarity of that feature. We introduce a 3 by 3 pixel operator which can be used to perform repeated erosion and dilation like operations that are functions of not only size and polarity, but also local derivatives or ratios of adjacent pixel values. These operations can be mixed to implement new types of nonlinear filtering functions. Examples are given in 1 and 2 dimensions.

In this report, we propose a new method of restoring an object from its cross sections having various contour shapes. The three-dimensional image of restored object is represented by its surface which consists of triangular patches. The object surface is smoothed according to requirement. This method reduces the complexity of triangulation algorithm and decrease the quantity of calculations to smooth surfaces.

Computer-assisted segmentation of vascular patterns in thermographic images provides the clinician with graphic outlines of thermally significant subcutaneous blood vessels. Segmentation strategies compared here consist of image smoothing protocols followed by thresholding and zero-crossing edge detectors. Median prefiltering followed by the Frei-Chen algorithm gave the most reproducible results, with an execution time of 143 seconds for 256 X 256 images. The Laplacian of Gaussian operator was not suitable due to streak artifacts in the thermographic imaging system. This computerized process may be adopted in a fast paced clinical environment to aid in the diagnosis and assessment of peripheral circulatory diseases, Raynaud's Disease3, phlebitis, varicose veins, as well as diseases of the autonomic nervous system. The same methodology may be applied to enhance the appearance of abnormal breast vascular patterns, and hence serve as an adjunct to mammography in the diagnosis of breast cancer. The automatically segmented vascular patterns, which have a hand drawn appearance, may also be used as a data reduction precursor to higher level pattern analysis and classification tasks.

A new vascular segmentation technique is introduced, which is based on a morphological image processing tool, dilation, and a modification of a recently developed class of image algorithms, snake transformations. Results are presented that enlighten the nature and properties of the technique, and applications are discussed.

This paper uses multiresolution representations in two new techniques for image enhancement and restoration. The first method, based on image pyramids, is used for approximating the convolution of an image with a given mask. In this technique, a filter is designed using a least squares (ls) procedure based on filter functions synthesized from the basic pyramid equivalent filters. This approximates the mask frequency characteristic. Next, enhancement involves linearly combining scaled and filtered pyramid levels, using weights obtained from the is procedure. By this method, filtering and pyramid image coding can be combined, efficiently integrating enhancement into the reconstruction procedure for the coded image. The second method is an adaptive noise reduction algorithm. An optimally filtered image is synthesized from the multiresolution levels, which in this case, are maintained at the original sampling density. Individual pixels of the image representation are linearly combined under a minimum mean square error criterion. This uses a local signal to noise ratio estimate to provide the best compromise between noise removal and resolution loss.

We have established a three-dimensional (3-D) imaging facility for reconstruction of serial two-dimensional (2-D) ultrasound images. In the facility, contiguous 2-D images are captured directly at the clinical site from the real-time video signals of a Labsonics serial ultrasound imager. The images are digitized and stored on an IBM PC. They are then transferred over an Ethernet communication network to the Image Processing Laboratory. Finally, the serial images are reformatted and the 3-D images are reconstructed on a Pixar image computer. The reconstruction method involves grey level remapping, slice interpolation, tissue classification, surface enhancement, illumination, projection, and display. We have demonstrated that 3-D ultra-sound images can be created which bring out features difficult to discern in 2-D ultrasound images.

A Bayesian generalized expectation - maximization (GEM) algorithm using a locally correlated Markov random field prior in the form of a Gibbs function is developed for emission tomography. A close-form coordinate gradient ascent M-step which updates the image pixels sequentially is derived. The resulting GEM Bayesian algorithm is applied to estimating the 3-D image parameters in the Poisson model of emission sources based upon simulation of a parallel collimated gamma camera.

Accurate and precise neuroanatomical descriptions are essential for the meaningful quantification of Positron Emission Tomographic (PET) images of the human brain. This task is difficult since the radio-tracer distribution imaged by PET does not neccessarily reflect structure. A two-dimensional brain atlas that features the use of a deformable Region-of-Interest (ROI) template has previously been shown to be an effective method for integrating anatomical (Magnetic Resonance Imaging - MRI) images and physiological (PET) data, with an observed reduction in the coefficient of variation amongst observers from 8.1% to 3.4% 1. This method has been adopted for routine MRI image correlation and PET image analysis at our centre. The atlas is most effective when used in studies in which PET and MRI images are acquired in planes parallel to those of the original brain atlas. To allow arbitrary image/atlas orientation, we have implemented a Volume of Interest (VOI) atlas that extends the ROI methodology to three dimensions. The major advantage is that, given adequate axial sampling, arbitrary orientations of PET and MRI scans can be matched to the VOI atlas with no loss in structural recovery. In addition, the VOI atlas can be employed in the quantification of metabolic parameters directly from volumetric PET studies rather than performing the analysis on a slice by slice basis.

The primary goal of any computer vision system is to construct scene descriptions similar to those formed by the human visual system. Biological image understanding is the automatic processing and interpretation of imagery of natural structures. The key constraint is the absence of explicit models of the objects anticipated to be in the scene. A clinician interprets imagery using principles of physiology and the image acquisition process. He learns to construct explicit models to recognize patterns from these principles, but is also able to infer an interpretation of an unknown structure from examples and physiology. Our overall goal is to develop a computer vision system to be used to interpret imagery of biological data using these principles. In this paper we present a paradigm for constructing image understanding systems designed to process natural images and present some applications of the system to hematology and radiography. The first step of our general approach is data acquisition. The second step is to find intrinsic properties at each pixel. The third step is to group pixels together to for homogeneous regions using the intrinsic properties. The final step is object recognition, where object type and location are determined from the quantitative features of the homogeneous regions. We shall outline the algorithms and data structures used at each step in the process, and their implementation. We shall describe in detail a key feature of this paradigm-forming homogeneous regions by geometry guided segmentation. The fact that classical algorithms do not always yield correct results led us to an algorithm that uses structural information and intensity values to group homogeneous regions. We shall illustrate this problem and the solution, as well as preliminary results of the application of the image understanding system.

Feldkamp has developed an approximate cone-beam tomography reconstruction algorithm. [1] This paper presents an analysis of some of the limitations of this algorithm and proposes a post-processing algorithm, which has been studied in simulations. Preliminary results suggest that by incorporating a priori constraints on the object reconstruction, the reconstruction volume within which satisfactory results are achieved can be expanded beyond that produced by the Feldkamp algorithm.

We expose a pretopological model based method for 3D reconstruction of surfaces. Pretopology provides concepts and results which enable us to reconstruct 3D surfaces betwen slices by mean of connectivity criteria and of what we call "addable points". First experimental results in the field of cardiac imaging are presented and discussed.

A sensitivity correction for a single turn receiver coil used to image a human knee is discussed. The coil consists of single band of copper conductor circling a plastic leg form and the patient's knee. While the excitation can be considered reasonably uniform across the field of view, the detection of the decaying MR signal by the single-turn solenoidal reciever coil is very non-uniform, falling off rapidly away from the copper conductor. We have modeled the sensitivity function of this coil as equivalent to that of a single-turn circular antenna with a known radius. A correction for this sensitivity variation may be approximated numerically and applied to the image in a post-processing manner. By a combination of qualitative and quantitative measures, the efficacy of this 'current-loop' correction scheme is measured by evaluating the uniformity of NiCl2 phantom images. A technique for applying this correction to clinical images of the knee is also discussed.

Recent impressive technological advances in imaging techniques for the human temporomandibular (tm) joint, and in enabling geometric algorithms have outpaced diagnostic analyses. The authors present a basis for systematic quantitative diagnoses that exploit the imaging advancements. A reference line, coordinate system, and transformations are described that are appropriate for tomography of the tm joint. These yield radiographic measurements (disk displacement) and observations (beaking of radiopaque dye and disk shape) that refine diagnostic classifications of anterior displacement of the condylar disk. The relevance of these techniques has been clinically confirmed. Additional geometric invariants and procedures are proposed for future clinical verification.

The clinical application of digital radiographic imaging has been the subject of several recent investigations (1-7). The digital radiographic image possesses certain advantages over the conventional analog radiographic image, such as wider dynamic range, potential for image manipulation, transfer, and storage, and radiation exposure reduction. However, the quality of image remains a valid concern to radiologists because they are accustomed to viewing conventional radiographs with a higher imaging quality.

A receiver operator characteristic (ROC) experiment designed to evaluate the diagnostic performance of 512, 1024 and 2048 monitors as compared to conventional analog films viewed on a lightbox was conducted. Fifty radiographs, 25 with and 25 without a simulated solitary nodule placed on the lung fields of an anthropomorphic chest phantom were used. These images were interpreted by six radiologists under the following six conditions: 512 and 1024 line resolution both with and without interactive contrast enhancement, and 2048 line resolution with contrast enhancement, and analog film on a conventional lightbox. As determined by the area under the ROC curve, which was the measure of diagnostic accuracy the 512 and 1024 line monitors performed below the analog film, the 2048 line monitor proved to be a viable alternative to the analog film.

This paper presents an efficient spiral scanning/sampling method and its associated reconstruction algorithm for echo planar Magnetic Resonance Imaging (MRI) systems. The spiral scan data are represented in terms of evenly spaced samples from a nonlinear two-dimensional transformation of the Cartesian spatial frequency domain. This transformation enables us to use unified Fourier reconstruction sampling principles [5] to obtain an accurate reconstruction based on a set of constraints imposed on the spiral parameters. The results are then utilized to develop efficient sampling strategies for spiral scans. Sampling efficiency for spiral data, analogous to the sampling efficiencies of hexagonally and rectangularly sampled data, is defined. A sampling scheme on a spiral is introduced that possesses a uniform sampling efficiency comparable to the sampling efficiency of rectangularly sampled data.

In this paper the concepts of pattern recognition, image processing and artificial intelligence are applied to the development of an intelligent cytoscreening system to differentiate the abnormal cytological objects from the normal ones in vaginal smears. To achieve this goal,work listed below are involved: 1. Enhancement of the microscopic images of the smears; 2. Elevation of the qualitative differentiation under the microscope by cytologists to a quantitative differentiation plateau on the epithelial cells, ciliated cells, vacuolated cells, foreign-body-giant cells, plasma cells, lymph cells, white blood cells, red blood cells, etc. These knowledges are to be inputted into our intelligent cyto-screening system to ameliorate machine differentiation; 3. Selection of a set of effective features to characterize the cytological objects onto various regions of the multiclustered by computer algorithms; and 4. Systematical summarization of the knowledge that a gynecologist has and the way he/she follows when dealing with a case.

Stereotactic surgery requires knowledge of cerebral structures derived from more than one image source. We have developed a PC-AT based workstation which accepts patient images, made with the stereotactic frame in place, from CT, MRI and DSA modalities. Reference markers on the frame are identified in the images to establish the coordinate geometry for each modality. Target points may be identified on each image type and trajectories of probe paths to these points defined. Targets identified on one set of images may be transferred automatically to other images of the same patient, in order, for example, to guarantee a vascular free path of approach to a target point deep within the brain. To date several hundred patients have had stereotactic surgery performed on the basis of plans using this system. Procedures included biopsy and aspiration of lesions, implantation of electrodes for the recording of deep EEG signals, and radiosurgical techniques based on the use of a high energy linear accelerator. We present clinical examples of the use of this system in typical stereotactic neurosurgery procedures, address stereoscopic applications, and discuss the results of inter-modality tests to establish the accuracy of the technique.

This paper presents a new technique, phase pseudocolor encoding, to realize medical electron microgragh pseudocolor processing. The density difference in the image can be changed into color difference by this method. Experimental results are given.

The paper describes and analyzes the Bell-Spline, a new algorithm that performs combined filtering and interpolation of a digital data set. In Digital Image Processing, the Bell-Spline algorithm can be effectively used to magnify or demagnify images of any integer or non-integer factor. Because of its inherent capability to filter the spatial frequencies of the image being interpolated, the algorithm can be tuned so as to prevent apparent loss of dynamic contrast of highly magnified images, or so as to minimize spatial aliasing of images that are demagnified or magnified by a small non-integer factor. The Bell-Spline algorithm performs the combined filtering and interpolation of an original two-dimensional pixel array with few operations per interpolated pixel. It is therefore suitable, and in fact has been used, for real-time image processing applications.

PET is the only imaging modality that provides doctors with early analytic and quantitative biochemical assessment and precise localization of pathology. In PET images, boundary information as well as local pixel intensity are both crucial for manual and/or automated feature tracing, extraction, and identification. Unfortunately, the present PET technology does not provide the necessary image quality from which such precise analytic and quantitative measurements can be made. PET images suffer from significantly high levels of radial noise present in the form of streaks caused by the inexactness of the models used in image reconstruction. In this paper, our objective is to model PET noise and remove it without altering dominant features in the image. The ultimate goal here is to enhance these dominant features to allow for automatic computer interpretation and classification of PET images by developing techniques that take into consideration PET signal characteristics, data collection, and data reconstruction. We have modeled the noise steaks in PET images in both rectangular and polar representations and have shown both analytically and through computer simulation that it exhibits consistent mapping patterns. A class of filters was designed and applied successfully. Visual inspection of the filtered images show clear enhancement over the original images.

The principle and technique of real-time pseudo-coloring enhancement for medical density images by liquid crystal light valve (LCLV) are discussed in detail. We present a simplied hybrid field effect mode of LCLV, and give the mathematical formula of projection light intensity. With this result, the characteristics of this pseudo-coloring method are discussed and the rule of optimum system is pointed out. Experiments prove our analysis method is effective. We set up a real-time pseudo-coloring system for medical density image, which can be used in practice. Some pseudo-coloring images of X-ray density pictures are shown in this paper.

A heuristic algorithm was designed and tested which allows for the computer generation of optimal preprocessing filter sequences for any class of image data. This method was employed in the search for an ideal technique of preprocessing echocardiographic image data prior to edge detection. The method appears potentially capable of producing efficient preprocessing sequences for this image class when compared to other widely used preprocessing methods. Furthermore, the search algorithm appears to be an acceptable way to reduce the infinitely large search space of convolution masks with various filtering properties.

We developed a digital data recorder (DDR) for realtime recording for use with a realtime digital cardio angiography system (DCA). The DDR can record/reproduce data at a rate of 8.7 MB/s. Also it can record data continuously for 20 min on a 1/2-inch high HC video cassette tape (data of 10.5 GB in total). Namely, it can realtime record images at a rate of 30 images in 512)(512 matrix/s or 7.5 images in 1024)(1024 matrix/s. Static images can be recorded in the majority mode with an improved reliability. Since the DDR can record data using commercially-available cassette tapes, the running costs are very low. Through clinical evaluation, We have a belief that the DCA using the DDR will take the place of the cine camera radiographic system.

The design of a parallel processor containing 60 32-bit processors with on-chip floating point hardware, 60 Mbytes of memory, hardware random number generation, and a high bandwidth, re-configurable interprocessor communications system is discussed. The system is especially well-suited to run Monte Carlo-based algorithms, such as simulated annealing, for optimization and estimation problems in nuclear medicine and other areas.

We have developed a rule-based, expert system for the analysis of digital angiographic images. The standard method of analyzing the sequence of images from an injection of contrast material ignores the information contained in the time sequence. This results in significant errors in the calculation of percent stenosis -particularly in vessels sloped with respect to the image plane. We have aimed at using the complete time sequence to improve the accuracy of detection and quantification of stenosis in arteries.

The aim of the paper is to present a segmentation method based on an extension to multi-gray-level images of the thresholding, named multithresholding and on a merging of so obtained regions using general knowledge concerning visual data. The system implemented firstly preprocesses the digitized images. Then, after a quadtree coding, detects the thresholds by gray-level histogram processing and obtains the regions by a labeling method. Finally the system performs the merging process, in which the detected regions are merged in order to refine the image segmentation. The merging is conceived as knowledge-based process oriented to perceptual concept utilization.

A PACS terminal contains a film digitizer which reads radiographic images digitally. Aliasing is often generated when the film digitizer reads images acquired on film using an antiscatter grid. Aliasing interferes with CRT diagnosis and therefore must be reduced. This paper clarifies the mechanism by which aliasing is generated and described how it can be reduced to a level acceptable for CRT diagnosis.