Morphological image processing (MORPH) is a useful technique for angiogram enhancement. We compare the technique with digital subtraction angiography (DSA) by viewing the filtering operation as the approximation of the mask image in DSA processing. Viewing the enhancement process in this manner allows us to judge the performance of the filter by measuring the RMS error between the DSA and MORPH images. By computing the error for a wide variety of structuring element (SE) shapes and sizes, we are able to pick the optimal SE as the one with minimum error. Two shapes of SEs are under investigation: ellipsoidal and an approximation to this, called the separable ellipsoid, obtained by dilating two one-dimensional elliptical SEs aligned in the row and column directions. The separable ellipsoid gives comparable results to those of the ellipsoid and allows a significant reduction in the computational time.

Recently, mathematical morphology has been used to develop efficient image analysis tools. This paper compares the performance of morphological and conventional edge detectors applied to radiological images. Two morphological edge detectors including the dilation residue found by subtracting the original signal from its dilation by a small structuring element, and the blur-minimization edge detector which is defined as the minimum of erosion and dilation residues of the blurred image version, are compared with the linear Laplacian and Sobel and the non-linear Robert edge detectors. Various structuring elements were used in this study: regular 2-dimensional, and 3-dimensional. We utilized two criterions for edge detector's performance classification: edge point connectivity and the sensitivity to the noise. CT/MR and chest radiograph images have been used as test data. Comparison results show that the blur-minimization edge detector, with a rolling ball-like structuring element outperforms other standard linear and nonlinear edge detectors. It is less noise sensitive, and performs the most closed contours.

An extension to the contrast-enhancement algorithm Adaptive Histogram Equalization (AHE) has been developed at our institution for use on digital chest images. Our algorithm, Artifact- Suppressed Adaptive Histogram Equalization (ASAHE), has been targeted for ultimate use on high-resolution radiological workstations. Recent modifications to the algorithm suggested that it would make nodules more easily detectable than the standard processing available with the computed radiography system at our disposal. We designed and carried out a carefully controlled observer experiment, detailed in this article, that used phantoms and simulated nodules to verify our hypothesis. The results showed that the ASAHE algorithm did produce a significant improvement in nodule detection.

The fully automated reporting of the extent of disease from coronary arteriograms is likely to be a four-step procedure: (1) segmentation of the center lines of the coronary tree from one or more angiographic projection(s); (2) detection of the arterial boundaries based on the center lines; (3) automated detection of possible narrowings in the coronary segments, and (4) identification of the coronary arteries. In this paper we will concentrate on the developments of techniques for the last step, the automated identification of coronary arterial segments. Our approach is based on the representation of the projection of the coronary tree by a graph so that graph matching techniques can be used for labeling of the coronary arterial segments. The coronary tree, which is a branching structure with possible crossings and overlaps as visualized in an angiographic projection, is assumed to be segmented more or less successfully from an angiogram with known projection geometry. A model graph is composed from the projection of a three-dimensional representation of the normal coronary tree, and matched with the data graph by inexact graph matching. Two types of graph representations will be discussed. In the first type arterial branching points are represented by nodes and arterial segments by arcs between nodes. In the second type the arterial segments are represented by nodes and relationships between the arteries by arcs. Nodes and arcs in both types of representation are attributed with a semantic vector of object (node) features or relational (arc) features, which is a mixture in the first type of representation. Both types have been implemented and we are currently in the process of determining optimal parameter values for the associated cost functions.

In order to improve the diagnosis of lung cancer from chest radiographs, we are developing a computerized scheme for the detection of pulmonary nodules. The computerized scheme is based on a difference-image technique, feature-extraction techniques and a combination of linear and nonlinear filtering methods. In order to improve the sensitivity of the nonlinear filtering method, we investigated various filter parameters. We found that with a combination of two filtering methods, one effective for subtle nodules and the other effective for moderately subtle nodules, sensitivity can be increased by approximately 15% at a given level of specificity.

The overall objective in neurosurgery is to localize and to treat a target volume within the cerebral medium as well as to understand its environment. To complete this objective, the 3D display of multimodality information is required; among them CT, MRI, angiography or atlas are particularly important. During the last decade solutions have been proposed to improve the rendering of 3D CT data sets. Applied to MRI without preprocessing these methods are not able to provide a good display quality for the brain anatomy for instance. This paper presents one year of experience in the 3D display of MRI volumes, oriented to the preparation of neurosurgery procedures (e.g. biopsy, epilepsy surgery): the main issues concerning the volume anisotropy, the brain segmentation and the volume rendering are explained. Emphasis is also given to the original way we propose to solve the brain segmentation problem by using automatic segmentation techniques (fuzzy masks and region valley following). The volume rendering technique is also presented and discussed (binary segmentation vs fuzzy segmentation). Finally, examples are presented concerning the use of 3D MRI images.

In order to help in analyzing PET data and really take advantage of their metabolic content, a system was designed and implemented to align and process data from various medical imaging modalities, particularly (but not only) for brain studies. Although this system is for now mostly used for anatomical localization, multi-modality ROIs and pharmaco-kinetic modeling, more multi-modality protocols will be implemented in the future, not only to help in PET reconstruction data correction and semi-automated ROI definition, but also for helping in improving diagnostic accuracy along with surgery and therapy planning.

A program called SCANNER (version 0.6) is described for performing 2-D interactive medical image segmentation using knowledge of anatomic shape. The knowledge is implemented in a radial contour model, which is a flexible, generic model that can accurately deform to fit the data, but which also encodes the expected shape and range of variation for a 2-D contour shape class. The model, which can describe contours that are single-valued distortions of a circle, is learned from training sets of similarly-shaped contours. Variation in the learned model allows it to provide search regions for low level edge detectors, thereby reducing the incidence of false edges. Initial evaluation of this system was performed for structures seen in 111 2-D CT images from 12 patients undergoing radiation treatment planning for cancer. The results suggest that the model is able to capture the cross-sectional expected shape and range of variation for several clinically-important structures (the liver, kidney, eye, and some tumors), that the knowledge-based approach should reduce the segmentation time over current manual methods by a factor between two and ten, and that the usefulness of the model decreases as variability of the structure increases.

Dynamic radiographic imaging of the mouth, larynx, pharynx, and esophagus during swallowing is used commonly in clinical diagnosis, treatment and research. Images are recorded on videotape and interpreted conventionally by visual perceptual methods, limited to specific measures in the time domain and binary decisions about the presence or absence of events. An image processing system using personal computer hardware and original software has been developed to facilitate measurement of temporal, spatial and temporospatial parameters. Digitized image sequences derived from videotape are manipulated and analyzed interactively. Animation is used to preserve context and increase efficiency of measurement. Filtering and enhancement functions heighten image clarity and contrast, improving visibility of details which are not apparent on videotape. Distortion effects and extraneous head and body motions are removed prior to analysis, and spatial scales are controlled to permit comparison among subjects. Effects of image processing on intra- and interjudge reliability and research applications are discussed.

Mapping of PET images with CT or MR has a very high potential for clinical applications. A crucial step in the mapping process is the registration of two sets of 3D image data. This work demonstrates a technique for real-time alignment of a 3D image set based on parallel processors and a high bandwidth data transporter. The rapid response afforded by this hardware makes it feasible to visually align two image sets by adjusting the angle and depth of cutting planes along a tilted axis in a trial and error manner.

We are developing various computer-vision schemes for the detection of masses and microcalcifications in digital mammograms. However, for the effective and efficient implementation of computer-aided diagnosis (CAD), appropriate man-machine interfaces must be developed. Thus, our plan is to incorporate our schemes into a dedicated workstation for use as a 'second opinion' in a mammographic screening program. Output from the computer would be displayed as an aid, leaving the final diagnostic decision with the radiologist.

We propose in this paper a method allowing to evaluate the overall 3D motion of the heart left ventricle (LV) from MRI data. Left ventricular motion was approximated through a linear model associated with an affine transformation to determine parameters for non-rigid motion of the LV. Representing the LV at each sampling time as a polyhedral surface and defining comparison of polyhedral surfaces provided a measurement of the model approximation error. The proposed method is applied to one normal and two diseased LV.

To estimate the high frequencies of a low resolution image, we propose a maximum likelihood solution whereby we incorporate a priori knowledge about the structure of the desired image. The algorithm was developed primarily for improving low resolution Magnetic Resonance spectroscopic images, where a priori structural information is obtained from anatomical features of the corresponding high resolution proton image. The algorithm involves a noniterative solution which can be easily regulated for noisy data sets.

Classification of tissue-types in Magnetic Resonance (MR) images has received considerable attention in the medical image processing literature. Interpretation of MR images is based on multiple images corresponding to the same anatomy. Relaxation Labeling (RL) is a commonly used low-level technique in computer vision. We present a solution method for the classification of tissue-types in brain MR images using RL. Information from multiple images is combined to form an initial classification. RL is used to resolve the ambiguity present in the initial classification by incorporating user specified compatibility coefficients. A problem with RL is the smoothing of borders between tissue-types. We include edge information from the original images to overcome this problem. This results in a marked improvement in performance. We present results from patient images.

In order to measure the accuracy of volume determinations of brain tissues from MRI images, a number of phantoms simulating these tissues were constructed and imaged in a clinical imager operating at 1.5 T. Method consisted of measuring signals from pure tissues and volume of interest and then solving the equations for partial volume. Comparisons of the results obtained from the images with the true volumes in the phantom yielded accuracies (rms) which ranged from 1.6% for a simple phantom to 6.7% for more complex but realistic phantoms.

Frequently MRI data does not cover completely the desired region of k space. We have investigated the reconstruction of MR images from incomplete data using the MEMSYS 3 maximum entropy algorithm. We compare conventional modulus images with the maximum entropy images. We show the importance of incorporating spatial correlation into the maximum entropy reconstructions in order to minimize truncation artifacts.

In this paper, we develop a new algorithm for the enhancement of low-contrast details of dental X-ray images using a two channel structure. The algorithm first decomposes an input image in the frequency domain into two parts by filtering: one containing the low frequency components and the other containing the high frequency components. Then these parts are enhanced separately using a transform magnitude modifier. Finally a contrast enhanced image is formed by combining these two processed parts. The performance of the proposed algorithm is illustrated through enhancement of dental X-ray images. The algorithm can be easily implemented on a personal computer.

A 12 bit CCD camera equipped with a linear sensor of 4096 photodiodes is used to digitize conventional mammographic films. An iso-precision conversion of the pixel values is performed to transform the image data to a scale on which the image noise is equal at each level. For this purpose film noise and digitization noise have been determined as a function of optical density and pixel size. It appears that only at high optical densities digitization noise is comparable to or larger than film noise. The quantization error caused by compression of images recorded with 12 bits per pixel to 8 bit images by an iso-precision conversion has been calculated as a function of the number of quantization levels. For mammograms digitized in a 40962 matrix the additional error caused by such a scale transform is only about 1.5 percent. An iso-precision scale transform can be advantageous when automated procedures for quantitative image analysis are developed. Especially when detection of signals in noise is aimed at, a constant noise level over the whole pixel value range is very convenient. This is demonstrated by applying local thresholding to detect small microcalcifications. Results are compared to those obtained by using logarithmic or linearized scales.

An automatic approach is proposed in this paper for the detection of left ventricular cavity boundaries from two dimensional echocardiograms. This object-oriented detection approach extracts the contour in exploring the information at different levels: local intensity information and geometrical features of the contour. A contour residual energy model is set up for this purpose and the geometrical features of the left ventricular cavity are introduced into contour detection through this model. The minimization of the contour residual energy can identify the cavity contours in very noisy environment.

This paper presents a quick and efficient way to detect and correct the linear and constant image-phase terms associated with MR images. We show that this correction provides us with the knowledge of the exact location of the DC term in k-space, which proves to be useful in the detection of x and y motion parameters. In addition, by displaying the real positive part of the image after the proposed correction, we can reduce background noise, motion artifacts and flow artifacts. Examples, analyses and results are provided to demonstrate the usefulness of the proposed detection and correction method.

In this paper, we present a comprehensive model for MR data acquisition in the presence of patient motion to provide a better understanding as to the source of motion artifacts. This model identifies and quantifies various sources of motion artifacts in 2-D Fourier imaging. We verify our model by comparing the results predicted by the model with actual MR images of phantoms subjected to motion with controlled parameters. We expect that the knowledge of the sources of artifacts will lead to new and better methods of compensating for them.

We propose a methodology for Positron Emission Tomography(PET) reconstruction using Multi Sensor Fusion techniques. Improvement of spatial resolution of PET images is achieved with the proposed method. Fusion of the PET images with the correlated MRI(CT) images is performed using fuzzy set theory and belief functions. Using the a priori values obtained from the fusion process, we reconstruct the image using Bayes' law. Simulation studies performed using the Shepp-Logan head phantom show the feasibility of the process and improvement in the reconstructed PET image.

A Bayesian method that incorporates a priori information derived from the spatially-correlated structural images in reconstruction of images acquired in positron emission tomography (PET) has been developed for the improvement of image quality. One source of a priori information that potentially can be very useful is the anatomic information extracted from the x-ray computed tomography (CT) and magnetic resonance (MR) images. For example, anatomic structures outlined on the correlated CT or MR images can be incorporated in the Bayesian method to aid in identifying boundaries in PET images. This new approach can improve the quality of the reconstructed PET images.

In order to evaluate the performance of our elastic matching system, we have created a digitized atlas from a young normal male brain, using 135 myelin-stained sections at 700 micron spacing. Software was written to enter and edit regional anatomic contours, which were stacked and aligned to create a three-dimensional atlas. This anatomy atlas can be elastically matched to CT or MRI scans, and then superimposed on an aligned set of PET scans of the same subject. We evaluated the matching system by comparing computer- generated contours with expert-defined contours for several subcortical structures, based on scans from six neurologically normal subjects. Previously, we reported the results from experiments that used CT data. This paper reports our progress towards the matching of the atlas with MRI scans.

We describe the implementation, experience and preliminary results obtained with a 3-D computerized brain atlas for topographical and functional analysis of brain sub-regions. A volume-of-interest (VOI) atlas was produced by manual contouring on 64 adjacent 2 mm-thick MRI slices to yield 60 brain structures in each hemisphere which could be adjusted, originally by global affine transformation or local interactive adjustments, to match individual MRI datasets. We have now added a non-linear deformation (warp) capability (Bookstein, 1989) into the procedure for fitting the atlas to the brain data. Specific target points are identified in both atlas and MRI spaces which define a continuous 3-D warp transformation that maps the atlas on to the individual brain image. The procedure was used to fit MRI brain image volumes from 16 young normal volunteers. Regional volume and positional variability were determined, the latter in such a way as to assess the extent to which previous linear models of brain anatomical variability fail to account for the true variation among normal individuals. Using a linear model for atlas deformation yielded 3-D fits of the MRI data which, when pooled across subjects and brain regions, left a residual mis-match of 6 - 7 mm as compared to the non-linear model. The results indicate a substantial component of morphometric variability is not accounted for by linear scaling. This has profound implications for applications which employ stereotactic coordinate systems which map individual brains into a common reference frame: quantitative neuroradiology, stereotactic neurosurgery and cognitive mapping of normal brain function with PET. In the latter case, the combination of a non-linear deformation algorithm would allow for accurate measurement of individual anatomic variations and the inclusion of such variations in inter-subject averaging methodologies used for cognitive mapping with PET.

Three-dimensional (3D) medical imaging deals with the visualization, manipulation, and measuring of objects in 3D medical images. So far, research efforts have concentrated primarily on visualization, using well-developed methods from computer graphics. Very little has been achieved in developing techniques for manipulating medical objects, or for extracting quantitative measurements from them beyond volume calculation (by counting voxels), and computing distances and angles between manually located surface points. A major reason for the slow pace in the development of manipulation and quantification methods lies with the limitations of current algorithms for constructing surfaces from 3D solid objects. We show that current surface construction algorithms either (a) do not construct valid surface descriptions of solid objects or (b) produce surface representations that are not particularly suitable for anything other than visualization. We present ALLIGATOR, a new surface construction algorithm that produces valid, topologically connected surface representations of solid objects. We have developed a modeling system based on the surface representations created by ALLIGATOR that is suitable for developing algorithms to visualize, manipulate, and quantify 3D medical objects. Using this modeling system we have developed a method for efficiently computing principle curvatures and directions on surfaces. These measurements form the basis for a new metric system being developed for morphometrics. The modeling system is also being used in the development of systems for quantitative pre-surgical planning and surgical augmentation.

We have developed an automated system to perform registration of medical images in three- dimensions (3-D), with accuracy on the order of the pixel sizes. This new approach attempts to integrate anatomical and/or functional information from images acquired by different modalities or at different times without manual operations. This method has been applied to registration of images acquired from X-ray computed tomography (CT), magnetic resonance imaging (MRI), and positron emission tomography (PET).

We have developed software capable of the three-dimensional tracking of objects in the brain volume, and the subsequent overlaying of an image of the object onto previously obtained MR or CT scans. This software has been developed for use with the Magnetic Stereotaxis System (MSS), also called the 'Video Tumor Fighter' (VTF). The software was written for a Sun 4/110 SPARC workstation with an ANDROX ICS-400 image processing card installed to manage this task. At present, the system uses input from two orthogonally-oriented, visible- light cameras and a simulated scene to determine the three-dimensional position of the object of interest. The coordinates are then transformed into MR or CT coordinates and an image of the object is displayed in the appropriate intersecting MR slice on a computer screen. This paper describes the tracking algorithm and discusses how it was implemented in software. The system's hardware is also described. The limitations of the present system are discussed and plans for incorporating bi-planar, x-ray fluoroscopy are presented.

The fast high-resolution X-ray CT scanner known as the Dynamic Spatial Reconstructor (DSR) has the facility for generating three-dimensional (3-D) coronary angiograms from a single-scan sequence during a nonselective injection of a single bolus of contrast medium. Such angiograms offer a method for visualizing the entire coronary arterial tree from any angle of view and give a possible means for detecting and quantitating coronary arterial stenoses. Previously, a skilled operator had to interact with a tedious workstation-based package to manually delineate the vascular tree and measure vessel cross-sectional areas. We have devised an automatic display-driven system for analyzing a 3-D DSR angiogram. The system consists of two separate menu-driven modules: (1) the Artery Extractor and (2) the Artery Display. The Artery Extractor performs automatic image-processing operations to extract the skeleton of the arterial tree. The Artery Display provides a graphical interface for interacting with the analyzed angiogram; using this interface, the operator can visualize structures in the angiogram and provide cues for making various measurements on the arterial tree.

We describe how adaptive cluster analysis and a linear model of tissue-mixing can achieve improved identification of tissues in MR images, with less reliance on human interaction. Our technique consists of two successive phases: a supervised training phase, which involves a small amount of human interaction; and an unsupervised training phase, which implements adaptive clustering. Two versions of unsupervised training are described. In the first version, which is comparable to earlier methods, no attempt is made to deal with the partial volume problem, whereas in the second version additional steps are taken to identify partial volume voxels and to estimate the tissue composition of such voxels. The reliability and accuracy of each of these versions are evaluated. We describe the results of comparative tests of our algorithms on a software phantom, MR images of a physical phantom, and in vivo MR images of human brains. These results indicate that accounting for partial volumes can improve the reliability of tissue identification.

The skeleton transformation is particularly useful in the field of image processing and may be computed by various techniques. This paper describes a new skeleton algorithm, which is based on the original concept of anchor point, i.e. of point that the skeleton is bound to contain. Its first step consists in extracting the desired anchor points. Then, the set X to skeletonize is progressively thinned in such a way that, on the one hand, the anchor points can never be removed and on the other hand, the homotopy of X cannot be modified. This operation is efficiently implemented thanks to a queue of pixels. The algorithm turns out to be extremely efficient and accurate on conventional computers. Furthermore, it allows to deal not only with standard skeletons, but also with such objects as minimal skeletons, homotopic markings, smoothed and pruned skeletons, etc. Its flexibility is also proved by the facts that it works both in the Euclidean and the geodesic cases, that its adaptation to any kind of grid is straight forward and that it can even be extend to n-dimensional images and to graphs skeletons in image analysis and shape recognition.

This paper presents a method for extracting the profile of the stomach by computational geometry. The stomach is difficult to recognize from an X-ray because of its elasticity. Global information of the stomach shape is required for recognition. The method has three steps. In the first step, the edge is enhanced, and then edge pieces are found as candidates for the border. Because the resulting border is almost always incomplete, a method for connecting the pieces is required. The second step uses computational geometry to create the global structure from the edge pieces. A Delaunay graph is drawn from the end points of the pieces. This enables us to decide which pieces are most likely to connect. The third step uses the shape of a stomach to find the best sequence of pieces. The knowledge is described in simple LISP functions. Because a Delaunay graph is planar, we can reduce the number of candidate pieces while searching for the most likely sequence. We applied this method to seven stomach pictures taken by the double contrast method and found the greater curvature in six cases. Enhancing the shape knowledge will increase the number of recognizable parts.

The cone beam X-ray transform modelizes the measurements on new 3D medical imaging devices using 2D detectors, for instance X-ray transmission tomographs using image intensifiers or gamma-ray emission tomographs using convergent collimators. The most commonly used reconstruction algorithm performs cone beam back projection (FELDKAMP 1984). But it induces some distortions for objects far from the plane of the cone apex. We have established an exact formula between the cone beam X-ray transform and the first derivative of the 3D Radon transform (GRANGEAT 1987). It shows that the distortions are induced by the shadow zone in the Radon domain related to the planes which intersect the object but not the apex trajectory. In the Radon domain, it becomes possible to restore the missing information by interpolation. Then the reconstruction principle is to compute and to invert the first derivative of the Radon transform. In this communication, we compare these two algorithms on reconstructions performed on simulated acquisitions. We study the shape and level distortions along lines parallel to the rotation axis. We present an analysis of the axial and radial variations of the Modulation Transfer Function (MTF) and of their distortions. We conclude that the Radon transform algorithm provides a regularized solution to the distortions, with optimized computing time on modern scientific computers.

The purpose of this paper is to present a fan beam short scan reconstruction technique for non- circular detector orbits. It applies a smoothing window to the projection data, and then a modified filtered backprojection method is performed. This technique is exact and the smoothing window is orbit independent; yet it requires more data than current circular-orbit short scan algorithms.

Fractal dimension is a measure of complexity of a fractal object. Its application in medical image analysis is widely accepted. In this paper we use 3-D surface tracking technique to calculate the fractal dimension of the brain. The fractal dimension in this case is a measure of the convolution of the cerebral surface. Series of MR images of the brain are read and interpolated into a 3-D volume such that each voxel is a cube. Each cube is equivalent to a box in the box-counting method. A surface tracking routine is applied to this 3-D volume. The number of faces (surface area) on the surface as well as volume inside the object can be obtained. Then we can change the voxel size and do the interpolation and surface tracking again. Based on these measurements for various voxel sizes we can calculate the fractal dimension. The fractal dimension measured for a brain specimen using the 3-D box counting technique is equal to 2.207.

We propose a new algorithm for registering 3D reconstructions of blood vessels from DSA with MR images of the brain. The registration transformation is determined by fitting the blood vessel tree reconstructed from DSA projections into the fissures of the brain derived from MR images, somewhat in the manner of fitting a key into a lock. The fit of the vessels into the fissures is guided by specific but simple anatomical knowledge. Preliminary evaluation of the algorithm has been carried out using data derived from a cadaver brain.

Transform coding has been used successfully for radiological image compression in the picture archival and communication system (PACS) and other applications. However, it suffers from the artifact known as 'blocking effect' due to division of subblocks, which is very undesirable in the clinical environment. In this paper, we propose a combined-transform coding (CTC) scheme to reduce this effect and achieve better subjective performance. In the combined- transform coding scheme, we first divide the image into two sets that have different correlation properties, namely the upper image set (UIS) and lower image set (LIS). The UIS contains the most significant information and more correlation, and the LIS contains the less significant information. The UIS is compressed noiselessly without dividing into blocks and the LIS is coded by conventional block transform coding. Since the correlation in UIS is largely reduced (without distortion), the inter-block correlation, and hence the 'blocking effect,' is significantly reduced. This paper first describes the proposed CTC scheme and investigates its information-theoretic properties. Then, computer simulation results for a class of AP view chest x-ray images are presented. The comparison between the CTC scheme and conventional Discrete Cosine Transform (DCT) and Discrete Walsh-Hadmad Transform (DWHT) is made to demonstrate the performance improvement of the proposed scheme. The advantages of the proposed CTC scheme also include (1) no ringing effect due to no error propagation across the boundary, (2) no additional computation and (3) the ability to hold distortion below a certain threshold. In addition, we found that the idea of combined-coding can also be used in noiseless coding, and slight improvement in the compression performance can also be achieved if used properly. Finally, we point out that this scheme has its advantages in medical image transmission over a noisy channel or the packet-switched network in case of channel error and packet loss.

In this paper, correlation properties of a class of chest X-ray medical images are examined and different 1-D and 2-D correlation models are applied to this class of image sources. Correlation structures for different scanning methods including row-by-row, column-by- column, diagonal and Peano are compared. It is suggested that the Peano scanning best reserves the inter-pixel correlation, which coincides with an earlier observation made by Lempel and Ziv. The rate-distortion properties are also discussed in terms of different 1-D and 2-D correlation models.

A class of Constraint Satisfaction Neural Networks (CSNN) is proposed for solving the problem of medical image segmentation which can be formulated as a Constraint Satisfaction Problem (CSP). A CSNN consists of a set of objects, a set of labels for each object, a collection of constraint relations linking the labels of neighboring objects, and a topological constraint describing the neighborhood relationships among various objects. Each label for a particular object indicates one possible interpretation for that object. The CSNN can be viewed as a collection of neurons that interconnect with each other. The connections and the topology of a CSNN are used to represent the constraints in a CSP. The mechanism of the neural network is to find a solution that satisfies all the constraints in order to achieve a global consistency. The final solution outlines segmented areas and simultaneously satisfies all the constraints. This technique has been applied to many images in different domains and the results show that this CSNN method is a very promising approach for image segmentation.

Dempster-Shafer (D-S) theory has been proposed to represent and propagate uncertain knowledge in expert systems. In this paper, we describe a medical image understanding system whose reasoning module employs the profound features of D-S theory such as compatible frames and multivariate belief functions. The proposed expert system, which is based on the blackboard architecture, is capable of mimicking the reasoning process of a human expert in dividing a set of correlated x-ray CT, T1- and T2-weighted MR images into semantically meaningful entities. In the blackboard-oriented system, different kinds of evidence provided by various knowledge sources form a hierarchy of evidential space to which D-S theory is applied. Several experimental results are given to illustrate the performance of the proposed system.

The automatic identification of precise left ventricular endocardial surfacing using echocardiography and cardiac MRI for the quantification of ejection fraction continues to be difficult. Standard image processing techniques have not been completely successful, principally because not all edge data directly corresponds to anatomical boundaries. Trained physicians must use considerable a priori information regarding normal human anatomy to `fill in' the missing details of a typical cardiac study. In this paper, we describe a new method to identify borders within medical images that incorporates an expert system based approach. Throughout the design of this approach, we maintained the following constraints: the system must easily capture expert information from multiple experts, over a variety of cardiac image formats, be trainable on any patient, and once trained provide fast execution times. The method was initially tested on echocardiographic images. Using a series of 2D echo image sequences, an expert traced endo- and epicardial edges in order to 'teach' the computer which pixels were myocardium and which were left ventricular cavity; each identified pixel was then convolved so as to amplify correlations found between the pixel and its neighbors. The result, applied prospectively at near real time speed, identified all pixels as being myocardium, left ventricular cavity or uncertain. As a consequence, endocardial borders were generated. These borders were then used to calculate systolic and diastolic areas and an area ejection fraction which proved to be within 2% of an expert traced and calculated area ejection fraction. These findings suggest this method holds promise for the capturing of expert knowledge of 2D cardiac ultrasound interpretation, and, through preliminary testing, we have shown its potential in the segmentation of 3D cardiac MRI volume sets for subsequent analysis and display.

Image analysis (IA) knowledge and scene analysis (SA) knowledge are cojointly used in Medical Image Interpretation (MII). Usually knowledge is implicitly incorporated into procedures, making the latter very application-dependent. On the contrary, we want to clearly separate them in the MII system we are designing. For this purpose, we selected the evaluation of the functional pulmonary fraction on SPECT images as a case example from which we characterized: (1) image analysis knowledge: -- the manipulated IA objects are identified and organized in a stable and minimal set of generalized IA objects. -- The used IA procedures are made generic to ensure the separation between SA and IA knowledge. We classify them according to their IA objects parameters. (2) domain-specific knowledge: -- control and ordering of such procedures according to scene information are identified and represented in an adequate form to be integrated in the system. The choices made above (theoretical definitions and organization of IA objects and procedures, and the separation between IA and domain-specific knowledge) that ensure generality and application- independence in MII is explained and their validity shown on the selected example. The approach is to be generalized to multimodality medical image interpretation.

There is a growing interest in using sophisticated knowledge-based systems for biomedical image interpretation. We present a principled attempt to use artificial intelligence methodologies in interpreting lateral skull x-ray images. Such radiographs are routinely used in cephalometric analysis to provide quantitative measurements useful to clinical orthodontists. Manual and interactive methods of analysis are known to be error prone and previous attempts to automate this analysis typically fail to capture the expertise and adaptability required to cope with the variability in biological structure and image quality. An integrated model-based system has been developed which makes use of a blackboard architecture and multiple knowledge sources. A model definition interface allows quantitative models, of feature appearance and location, to be built from examples as well as more qualitative modelling constructs. Visual task definition and blackboard control modules allow task-specific knowledge sources to act on information available to the blackboard in a hypothesise and test reasoning cycle. Further knowledge-based modules include object selection, location hypothesis, intelligent segmentation, and constraint propagation systems. Alternative solutions to given tasks are permitted.

Recently, the problem of automatic left ventricular boundary detection has yielded to methods that rely on the matching of the image data to a model of the heart. The problem of matching such a model to data is further complicated by the fact that heart position, orientation and size can vary widely for various patient studies. In this paper, we propose a method for generation of the left ventricular boundary model for both normal or abnormal (interior and inferior) shapes based on a general notion of the average shape, size and orientation of the left ventricular boundary obtained from a clinical data set with 30 patients in each category. The general description is invariant to position and contains the mean and limits of the size and orientation of the heart boundaries as well as descriptors for different shapes. A total of 100 frames/cycle is assumed for a global model with the ES frame being at frame 50. The descriptor of each frame of the 100 samples is obtained by interpolation between adjacent original frames. A particular model for a patient cycle with known number of frames and known position of ED and ES is carried on by an interpolation over the global model. This general model is employed for the partial matching against boundary fragments derived from Canny edge detection. The mean square error between the slope angle (shape descriptor) of the model and that of the boundary provides a cost function. The minimum of the function, corresponding to the best match, is obtained by varying the size of the model and the relative offset between the beginning of the model and the boundary fragment. After the parameters of the model are determined, it can be placed in the x-y plane to select more boundary fragments from the edge image, or it can be used for ranking the boundary fragments or complete boundaries. This matching is specially useful for open contours with preferences on their size and/or on their orientation, or even bounds on those values.

A microcomputer-based image processing system is used to digitize and process serial sections of CT/MRI scan and reconstruct three-dimensional images of brain structures and brain lesions. The images grabbed also serve as templates and different vital regions with different risk values are also traced out for 3D reconstruction. A knowledge-based system employing rule-based programming has been built to help identifying brain lesions and to help planning trajectory for operations. The volumes of the lesions are also automatically determined. Such system is very useful for medical skills archival, tumor size monitoring, survival and outcome forecasting, and consistent neurosurgical planning.

Intensity interpolation is an operation that recovers the intensity information in the gaps of serial cross sections of three-dimensional objects. It is very important in object reconstruction, especially when the reconstructed object is to be further manipulated for visualization, dissection, or slicing from arbitrary angles. One of the most difficult problems in reconstruction is how to handle the branching situations. Branching occurs where one region in one slice has multiple corresponding regions in the other slice. In this paper, a method to handle branching in intensity interpolation from serial cross sections is proposed. It consists of eight major steps: (1) partial contour to contour correspondence establishment, (2) displacement field determination, (3) intermediate contour generation, (4) interior rim generation, (5) rim correspondence establishment, (6) line segment correspondence establishment, (7) intensity determination, and (8) gap filling. As a significant extension of our previous work which performs intensity interpolation for one-to-one case, the proposed method contains a novel technique to form a displacement field which is used in generating intermediate contours (steps 1 to 3), and adapts some of the previous techniques (steps 4 to 8) to select appropriate corresponding regions for intensity interpolation.

Digital chest radiographs are often too bright and/or lack contrast when viewed on a video display. This is often seen when the radiographs are taken of patients with dense lungs, have incorrect X-ray exposure technique or have inappropriate image pre-processing performed by the image acquisition system (CR system or laser scanner). This paper describes a method which can automatically provide brightness and contrast adjustments to selectively enhance either soft or dense tissues. This reduces viewer interaction and improves displayed image quality. An algorithm has been developed to analyze the gray-level histogram of a chest radiograph and determine the breakpoints that separate: (1) the region outside the patient (background), (2) the radiographically soft tissues, and (3) the radiographically dense tissues. From these breakpoint values, a series of piecewise linear look-up tables (LUTs) are generated to selectively enhance either the soft tissues or the dense tissues. This is performed by (1) varying the contrast in the patient background to achieve the desired overall brightness, (2) selectively increasing the contrast of the tissue region of interest, and (3) reducing or maintaining the contrast of the remaining region. The resulting LUTs are applied to the original image via video display.

The efficacy of diagnostic thermal imaging, the visualization of abnormal distribution of temperature over the human skin, can be significantly augmented by computerized image processing procedures that overcome the limitations of subjective image assessment. This paper reviews diagnostic thermal imaging and describes common image processing approaches applicable to the analysis of static thermal images and of time series of images that provide diagnostic information about the dynamics of neurological regulation of skin temperature.

A current goal in medical image processing is image segmentation and texture analysis. Although orthogonal transforms are interesting for image analysis they do not preserve spatial localization. In this paper we propose and compare some tools for local spectrum analysis. The first uses an evolutive windowed Fourier or Cosine Transform while the second is obtained via the 2D Wigner Distribution. Both of these transforms result in a collection of 2D spectra. In order to make the comparison easier we extract some parameters from each local spectrum. At last the first results obtained on a textural image and a medical image are presented.

In this paper we describe a novel noise smoothing method based on a nonparametric statistic runs test. We assume that the data bits of a pixel can be divided into signal bits and noise bits. The signal comprises the most significant bits and the noise bits are the least significant ones. The idea in this smoothing method is to preserve the signal bits and only modify the noise bits. The number of noise bits of each pixel is determined based on the runs in the neighborhood. If the number of noise bits is zero then no smoothing is necessary. The degree of smoothing is a function of the number of noise bits. Using this technique we are able to smooth only noisy areas without reducing the spatial resolution in the image. The algorithm is easy to implement. The application of the smoothing algorithm on a chest image was given.

Computer based image analysis offers a wide range of techniques which can be used to make objective and reproducible detection and measurement of subvisual features of microscopic images of cells. Texture information is understood to be quite relevant to cyto-diagnosis, and is badly in need for quantification. We describe here two alternatives for texture analysis; the frequency spectrum and the runlength methods. Some experiments with both simulated and real cell image data have been carried out using these methods and preliminary results obtained. It is shown that these results are promising in that they seem to correlate with some medical expectations.

A method is presented to automatically determine the three-dimensional geometry of a stationary vessel tree from orthogonal biplanar digital angiographic image sequences, without a priori knowledge or user interaction. Vessels are identified unambiguously by their position in each projection and by comparison of time-density curves. Single-plane angiographic sequences are used to illustrate the technique, simulating a single slice of three-dimensional data. Projections are taken in each direction, yielding one row and one column of data which are used to reconstruct the vessel geometry. It is demonstrated that overlapping vascular regions can be resolved by temporal processing of angiographic image sequences, although high-quality reconstruction has not yet been achieved.

The signal intensity in a magnetic resonance image is not only a function of imaging parameters but also of several intrinsic tissue properties. Therefore, unlike other medical imaging modalities, magnetic resonance imaging (MRI) allows the imaging scientist to locate pathology using multispectral image segmentation. Multispectral image segmentation works best when orthogonal spectral regions are employed. In MRI, possible spectral regions are spin density (rho) , spin-lattice relaxation time T1, spin-spin relaxation time T2, and texture for each nucleus type and chemical shift. This study examines the ability of multispectral image segmentation to locate breast pathology using the total hydrogen T1, T2, and (rho) . The preliminary results indicate that our technique can locate cysts and fibroadenoma breast lesions with a minimum number of false-positives and false-negatives. Results, T1, T2, and (rho) algorithms, and segmentation techniques are presented.

Three-dimensional image registration is computationally intensive and therefore slow. Parallelizing the primitive computation processes involved in registration can decrease the time. Our approach for improving image processing speed is to utilize a massively parallel single instruction stream -- multiple data stream (SIMD) computer. This paper presents the performance results comparing a massively parallel architecture to a four processor vector/parallel system and a conventional serial computer.

In this paper we report on an experiment to compare the relevance of various image enhancement methods for improving the visibility of pathologies on digitized chest radiographs. The five pathologies tested in our trial are pulmonary nodules, air bronchograms, paratracheal abnormalities, pneumothoraces, interstitial lung diseases. The first three are examples of situations where focus is put on shape, borders and content of the pathology, the next is an example of situations where the visualization of a subtle line is required and the last one is an example of diffuse disease where the perceivability of details is important. Eight image enhancements were tested and included both pixel based gray-level transformation such as, windowing, statistical differencing, polynomial transform, histogram equalization, histogram hyperbolization, and spatial enhancement such as, unsharp masking with different masks and a Sobel detector. For each pathology we recommend two or three acceptable transformations.

This paper describes the use of magnetic resonance imaging to produce quantitative information about heart motion. Motion estimation techniques utilizing convex set projections and simple block shift matching are applied to determine direction and magnitude of heart motion. These methods are tested and compared using simulated motion on a single MRI frame and actual MRI cine data. The simulated motion is in the form of translational only and linear velocity fields. Motion estimation is performed using image pairs with and without additive noise. Results are given which quantify the effectiveness of the algorithms.

The accurate measurement of kidney glomerular basement membrane (GBM) width from transmission electron micrographs provides vital information in the diagnosis of some kidney diseases. A novel and effective algorithm has been developed for the semi-automatic detection of the GBM by image analysis techniques. The method makes use of the gradual variation of the local features that are characteristic of the membrane. Starting from a seed point specified by the user, local features within a small window are computed to give a feature score. Feature scores for adjacent neighborhoods are also determined, and windows that satisfy a similarity criterion are linked to produce the centerline of the membrane. A region-growing process completes the segmentation procedure.

A class of artificial intelligence (Al) programs known as neural networks are well suited to pattern recognition. A neural network is trained rather than programmed to recognize patterns. This differs from "expert system" Al programs in that it is not following an extensive set of rules determined by the programmer, but rather bases its decision on a gestalt interpretation of the image. The "bullseye" images from cardiac stress thallium tests performed on 50 male patients, as well as several simulated images were used to train the network. The network was able to accurately classify all patients in the training set. The network was then tested against 50 unknown patients and was able to correctly categorize 77% of the areas of ischemia and 92% of the areas of infarction. While not yet matching the ability of a trained physician, the neural network shows great promise in this area and has potential application in other areas of medical imaging.

Human vision, for the purpose of shape recognition, decomposes three-dimensional shapes into subunits. Simple rules, stated in the language of differential geometry, can describe this decomposition.