To avoid missing breast malignancies, large numbers of benign breast lesions must be biopsied. To reduce the number of benign biopsies, a computer aided diagnosis (CAD) method has been developed which is based on tissue texture surrounding calcifications. When core samples are obtained stereotaxically, a digital record of the area biopsied is available. This method has been tested on 82 biopsies containing calcifications. Of these, 52 were benign and 30 were malignant. A region of interest centered on the biopsied area was processed to obtain texture features. Both co-occurrence and fractal features were collected and used with stepwise linear discriminant analysis to isolate useful features. A jackknife method identified ten features that gave a probability distribution associated with malignancy. Because of this association, a probability could be selected which eliminated 12 of the 52 benign biopsies without missing a malignancy. Thirty-nine could be avoided if five malignancies could be followed rather than biopsied. Unfortunately, four of these five missed malignancies do not have strong visual signs of malignancy and so the texture measurement error would not be overruled by radiological signs.

This paper describes an automatic computer searching system for detecting clustered microcalcifications. A fuzzy classification modeling was employed to extract each suspected microcalcification possessing similar physical parameters. Therefore, only those possible classes were evaluated using a sophisticated convolution neural network which requires a great deal of computation and serves as a discriminator. Based on the detected spots, many of them are true microcalcifications, the computer can easily make a determination when 5 spots are located within a defined region. However, when a cluster consists of only two to four suspicious spots a fuzzy function was used to determine the inclusion of other spots near the cluster. This can be very important for the detection of subtle cases. The membership of the latter fuzzy function was composed of the distance between the suspected spots as well as the output values of the convolution neural network. We have tested the improved algorithms on our research database consisting of 45 mammograms. The results indicated that the fuzzy classification modeling decreased the number of false-positives from 2,874 to 1,067 suspected spots per image without increasing any false-negative detection. The over-all performance in the detection of clustered microcalcifications through the updated algorithms was 90% sensitivity at 0.5 false-positive per image. The computation time using a DEC-Alpha workstation was decreased from 5 minutes to about 3 minutes per image.

A multi-stage system with image processing and artificial neural techniques is developed for detection of microcalcification in digital mammogram images. The system consists of (1) preprocessing stage employing box-rim filtering and global thresholding to enhance object-to- background contrast; (2) preliminary selection stage involving body-part identification, morphological erosion, connected component analysis, and suspect region segmentation to select potential microcalcification candidates; and (3) neural network-based pattern classification stage including feature map extraction, pattern recognition neural network processing, and decision-making neural network architecture for accurate determination of true and false positive microcalcification clusters. Microcalcification suspects are captured and stored in 32 by 32 image blocks, after the first two processing stages. A set of radially sampled pixel values is utilized as the feature map to train the neural nets in order to avoid lengthy training time as well as insufficient representation. The first pattern recognition network is trained to recognize true microcalcification and four categories of false positive regions whereas the second decision network is developed to reduce the detection of false positives, hence to increase the detection accuracy. Experimental results show that this system is able to identify true cluster at an accuracy of 93% with 2.9 false positive microcalcifications per image.

Computer-aided diagnosis can be defined as a diagnosis made by a radiologist who takes into account the computerized analysis of the radiograph. The potential advantage of this approach is that errors by human observers that are caused by intra-observer variation, can be caught by the computer, which performs a thorough, methodical search of the image. However, if the input image to computer comes from a digitized radiograph, then the process of digitization introduces variability into the computer analysis of the image. Because the film digitizer samples the image, re-digitization will not necessarily produce the exact same digital image, and subsequently, the computerized scheme may produce different results on the different samples of the original image. The purpose of this study was to examine the amount of variability introduced by film digitization by measuring the reproducibility of our computerized scheme for the detection of clustered microcalcifications on mammograms when the mammogram is digitized a number of times. We have found in general that if the detected cluster (either a true or false cluster) corresponds to some obvious anatomical structure, then the cluster will be re-detected upon re-digitization. If, however, the cluster corresponds to low contrast objects within the breast, including image noise, then the cluster is not likely to be re- detected. As a result, true clusters tend to be much more reproducible than false-positive clusters. In the future, when direct digital radiographs are used as input to the computerized schemes, problems with reproducibility will not exist.

A database of 145 mammograms containing biopsy proven malignant or benign microcalcifications was digitized with a laser scanner at a pixel size of 35 micrometer by 35 micrometers. Digitization at larger pixel sizes was simulated by averaging adjacent pixels. The individual microcalcifications were segmented from the digital images with region growing and adaptive gray level thresholding techniques. The characteristics of the individual microcalcifications were analyzed with visibility descriptors and shape descriptors. The variations of visibility and shape of the microcalcifications in a cluster were evaluated by the standard deviation, the coefficient of variation, and the maximum of each of the descriptors. In addition, texture features were extracted from the spatial gray level dependence (SGLD) matrices in the region containing the cluster of microcalcifications. A genetic algorithm (GA) was used to select features from the multidimensional morphological and texture feature space. Linear discriminant analysis was employed to classify the microcalcifications as benign or malignant based on the selected features using a cross-validation training and test method. The classification accuracy was evaluated by the area, A_z, under the receiver operating characteristic (ROC) curve. We varied the pixel size from 35 micrometers by 35 micrometers to 140 micrometers by 140 micrometers. The classification accuracy using morphological features decreased as the pixel size increased from 35 micrometers by 35 micrometers. In the texture feature space, the classification accuracy did not depend strongly on the pixel size. In the combined morphological and texture feature space, the classification accuracy also tended to decrease as the pixel size increased from 35 micrometer by 35 micrometer. The effect of spatial resolution on computerized classification depends on the features extracted for the classification task. We have not explored all possible features or feature parameters in this study and therefore the high spatial resolution information in the images may not have been fully exploited. Further studies to develop new features are being conducted and the impact of the spatial resolution needs to be further analyzed with a larger data set.

A new rubber-band straightening transform (RBST) and texture analysis were developed for classification of masses on mammograms as malignant or benign. A database of 168 mammograms containing biopsy-proven malignant and benign breast masses was digitized at a pixel size of 100 micrometer by 100 micrometer. Regions of interest (ROIs) containing a single mass from each mammogram were extracted by an experienced radiologist. A clustering algorithm was employed for automatic segmentation of each ROI into a mass object and background tissue. A 40-pixel-wide region of the gray-scale image surrounding the mass was transformed onto the Cartesian plane using RBST. The first row of the transformed image contained the pixels along the mass edge. The j^th row of the transformed image contained the pixel values at a distance of j pixels from the edge in the normal direction to the edge contour. In the RBST image, mass margins appeared as approximately horizontal edges, and spiculations appeared as approximately vertical lines. Thus, edge sharpness and possible spiculations could be better characterized with directional texture measures. Texture features were extracted from spatial gray-level dependence matrices calculated from the RBST images. Linear discriminant analysis was used for classification, and receiver operating characteristic (ROC) analysis was used to evaluate the classification accuracy. The area A_z under the test ROC curve reached 0.87 using cross-validation and 0.89 using the leave-one-case-out method.

In this investigation we explore genetic algorithms as a technique to train the weights in a feed forward neural network designed to predict breast cancer based on mammographic findings and patient history. Mammograms were obtained from 206 patients who obtained breast biopsies. Mammographic findings were recorded by radiologists for each patient. In addition, the outcome of the biopsy was recorded. Of the 206 cases, 73 were malignant while 133 were benign at the time of biopsy. A genetic algorithm (GA) was developed to adjust the weights of an artificial neural network (ANN) so that the ANN would output the outcome of the biopsy when the mammographic findings were given as inputs. The GA is a technique for function optimization that reflects biological genetic evolution. The ANN was a fully connected feed- forward network using a sigmoid activation with 11 inputs, one hidden layer with 10 nodes, and one output node (benign/malignant). The GA approach allows much flexibility in selecting the function to be optimized. In this work both mean-squared error (MSE) and receiver operating characteristic (ROC) curve area (Az) were explored as optimization criteria. The system was trained using a bootstrap sampling. Optimizing for the two criteria result in different solutions. The 'best' solution was obtained by minimizing a linear combination of MSE and (1-Az). ROC areas were 0.82 plus or minus 0.07, somewhat less than those obtained using backpropagation for ANN training: 0.90 plus or minus 0.05. This is the first description of a genetic algorithm for breast cancer diagnosis. The novel advantage of this technique is the ability to optimize the system for maximizing ROC area rather than minimizing mean squared error. A new technique for computer-aided diagnosis of breast cancer has been explored. The flexibility of the GA approach allows optimization of cost functions that have relevance to breast cancer prediction.

We have designed and implemented a circumscribed lesion detection algorithm, based on the Hough Transform, which will detect zero or more approximately circular structures in a mammogram over a range of radii from a few pixels to nearly the size of the breast. We address the geometrical behavior of peaks in Hough parameter space {x, y, r} for both the true radius of a circular structure in the image (r equals r_o), and for the parameter r as it passes through this radius. In addition, we scale the peaks in Hough parameter space by re-analyzing the contrast of the region in the mammogram underlying the circular disk indicated by the peak. These scaled peaks are mapped to circular disks and accumulated in an output feature image. This process defines a continuously scaled pixel level output. That output, a local image measurement or 'feature,' suggests the likelihood that a pixel is located inside a circular structure, irrespective of the radius of the structure and overall mammogram contrast. These feature values are evaluated by fast qualitative and quantitative performance metrics which permit circumscribed lesion detection features to be initially evaluated without a full end-to-end classification experiment.

The purpose of this work is to develop and to evaluate an automatic analysis technique for quantitative assessment of cardiac function from cine MRI and to identify regional alterations in synchronicity based on Fourier analysis of ventricular wall motion (WM). A temporal analysis technique of left ventricular wall displacement was developed for quantitative analysis of temporal delays in wall motion and applied to gated cine 'dark blood' cardiac MRI. This imaging technique allows the user to saturate the blood both above and below the imaging slice simultaneously by using a specially designed rf presaturation pulse. The acquisition parameters are: TR equals 25 - 60 msec, TE equals 5 - 7 msec, 0 equals 25 degrees, slice thickness equals 10 mm, 16 to 32 frames/cycle. Automatic edge detection was used to outline the ventricular cavities on all frames of a cardiac cycle. Two different segmentation techniques were applied to all studies and lead to similar results. Further improvement in edge detection accuracy was achieved by temporal interpolation of individual contours on each image of the cardiac cycle. Radial analysis of the ventricular wall motion was then performed along 64 radii drawn from the center of the ventricular cavity. The first harmonic of the Fourier transform of each radial motion curve is calculated. The phase of the fundamental Fourier component is used as an index of synchrony (delay) of regional wall motion. Results are displayed in color-coded maps of regional alterations in the amplitude and synchrony of wall motion. The temporal delays measured from individual segments are evaluated through a histogram of phase distribution, where the width of the main peak is used as an index of overall synchrony of wall motion. The variability of this technique was validated in 10 normal volunteers and was used to identify regions with asynchronous WM in 15 patients with documented CAD. The standard deviation (SD) of phase distribution measured in short axis views was calculated and used to identify regions with asynchronous wall motion in patients with coronary artery disease. Results suggest that this technique is more sensitive than global functional parameters such as ejection fraction for the detection of ventricular dysfunction. Color coded parametric display offers a more convenient way for the identification and localization of regional wall motion asynchrony. Data obtained from endocardial wall motion analysis were not significantly different from wall thickening measurements. The innovative approach of evaluating the temporal behavior of regional wall motion anomalies is expected to provide clinically relevant data about subtle alteration that cannot be detected through simple analysis of the extent (amplitude) of wall motion or myocardial thickening. Temporal analysis of regional WM abnormality from cine MRI offers an innovative and promising means for objective quantitative evaluation of subtle regional abnormalities. Color coded parametric maps allowed a better identification and localization of regional WM asynchrony.

Quantification of the severity of the multiple sclerosis (MS) disease through estimation of lesion volume via MR imaging is vital for understanding and monitoring the disease and its treatment. This paper presents a novel methodology and a system that can be routinely used for segmenting and estimating the volume of MS lesions via dual-echo spin-echo MR imagery. An operator indicates a few points in the images by pointing to the white matter, the gray matter, and the CSF. Each of these objects is then detected as a fuzzy connected set. The holes in the union of these objects correspond to potential lesion sites which are utilized to detect each potential lesion as a fuzzy connected object. These 3D objects are presented to the operator who indicates acceptance/rejection through the click of a mouse button. The volume of accepted lesions is then computed and output. Based on several evaluation studies and over 300 3D data sets that were processed, we conclude that the methodology is highly reliable and consistent, with a coefficient of variation (due to subjective operator actions) of less than 1.0% for volume.

Single-frame quantitative coronary angiography (QCA) requires a human operator to select one frame for analysis out of typically 90 or more acquired in each sequence. QCA measurements for the same arterial segment differ from frame to frame due to a variety of factors including: rapid coronary motion, overlapping arteries, patient scatter, and structure noise from imaging the ribs, spine and diaphragm. These factors make frame selection a source of diagnostic variability. Test images were generated by radiographically projecting 3- D model arteries of known diameters onto patient coronary angiograms. Our objective is to eliminate subjective selection of a single frame and improve the overall accuracy of QCA by utilizing all of the image frames in a sequence. Diameters of the artery within each image were measured with an automated edge detection algorithm. Our approach to improve the accuracy of the diameter measure is to search for the coronary artery edge in both the spatial (individual image frame) and temporal domain (all frames in the sequence that show the edge).

A method has been developed for in-room computer reconstruction of the three-dimensional (3-D) coronary arterial tree from routine biplane angiograms acquired at arbitrary angles and without using calibration objects. The proposed method consists of four major steps: (1) segmentation of vessel centerlines and bifurcation points and measurement of vessel diameters in coronary angiograms, (2) determination of biplane imaging parameters in terms of a rotation matrix R and a translation vector t based on the identified bifurcation points, (3) recovery of 3-D coronary arterial tree based on the calculated biplane imaging parameters, correspondences of vessel centerlines, and diameters, and (4) rendering of reconstructed 3-D coronary tree and estimation of optimal view of selected arterial segments. Angiograms from fifteen patients were utilized for 3-D reconstruction for each patient's coronary arterial tree. The biplane imaging geometry was first determined without a calibration object, and the 3-D coronary arterial trees were reconstructed including both left and right coronary artery systems. Various 2-D projection images of the reconstructed 3-D coronary arterial tree were generated and compared to other viewing angles obtained in the actual patient study. Similarity between the real and reconstructed arterial structures was excellent. The accuracy of this method was evaluated by using a computer-simulated coronary arterial tree. Root-mean-square (RMS) errors in the 3-D position and 3-D configuration of vessel centerlines and in the angles defining the R matrix and $t vector were 1.2 - 5.5 mm, 0.3 - 2.0 mm, and less than 1.5 and 2.0 degrees, respectively, when using 2-D vessel centerlines with RMS normally distributed errors varying from 0.7 - 4.2 pixels (0.25 - 1.26 mm).

A method for 3D cone beam reconstruction of cerebral vasculature (both morphology and grayscale) from a limited number (less than 10) of digital subtraction angiographic (DSA) projections obtained with a standard biplane C-arm x-ray system is described. The reconstruction method includes geometric calibration of the source and detector orientation, spatial image distortion correction, and algebraic reconstruction technique (ART) with non- negativity constraint. Accuracy of voxel gray scale values estimated by ART is enhanced by determination of weights based on the intersection volume between a pyramidal ray and cubic voxel. The reconstruction is accelerated by retaining only the vessel containing voxels and distributed computing. Reconstruction of a phantom containing fiducial markers at known 3D locations demonstrated that the reconstructed geometry is accurate to less than a pixel width. Reconstruction is also obtained from an anatomic skull phantom with an embedded cerebral vasculature reproduction that includes an aneurysm. Three dimensional reconstruction exhibited the necessary details, both structural and grayscale.

We are developing simple and efficient tools for assessing qualitative and quantitative information on 3D carotid anatomy and velocity distributions by integrating 2D color Doppler and B-mode images. Transverse color Doppler images of the carotid bifurcation were obtained during relatively uniform, manual scanning of the transducer. Longitudinal views were also recorded. The transverse images obtained at or near systole were automatically extracted by analysis of the flow data. The flow regions in the lumen in the individual images were segmented using the interface between the flow colored and non-colored, i.e., gray level, regions in the systolic images. A 3D volume was generated by stacking and interpolating each component. Representative longitudinal projections were then obtained by resampling this 3D volume according to the vessel centerline and integrating longitudinal slices of this volume to take into account the transducer resolution. The computed longitudinal images were matched to the longitudinally acquired images using a Procrustes technique, and a registered 3D reconstruction of the flow patterns, lumen and intima was obtained. Preliminary results show that the computed longitudinal projections of the 3D reconstructed volume are globally consistent with the acquired scans.

We propose a stopping criterion for the EM-MLE algorithm for PET, based on the properties of the multiplicative correction factors (Delta) x applied to the image vector after each iteration. The criterion is based on the relationship between the image quality descriptors and the statistical properties of the updating coefficients. Our study was performed with Monte Carlo-generated projections initially for an ideal scanner and noise-free case, in order to study the properties of the algorithm. It was then extended to include realistic data from the Hoffman brain phantom, by detailed Monte Carlo simulation of scattering, absorption, random and multiple coincidences, including a Monte Carlo simulation of the transmission scan for attenuation correction. The criterion has been applied to reconstructions from real data from the CTI-831 NeuroPET scanner. The stopping criterion as well as results from reconstructions of measured projection data are discussed.

Geometric models have received increasing attention in medical imaging for tasks such as segmentation, reconstruction, restoration, and registration. In order to determine the best configuration of the geometric model in the context of any of these tasks, one needs to perform a difficult global optimization of an 'energy' function that may have many local minima. Explicit models of geometry, also called deformable models, snakes, or active contours, have been used extensively to solve image segmentation problems in a non-Bayesian framework. Researchers have seen empirically that multi-scale analysis is useful for convergence to a configuration that is near the global minimum. In this type of analysis, the image data are convolved with blur functions of increasing resolution, and an 'optimal' configuration of the snake is found for each blurred image. The configuration obtained using the highest resolution blur is used as the solution to the global optimization problem. In this article, we use explicit models of geometry for a variety of Bayesian estimation problems, including image segmentation, reconstruction and restoration. We introduce a multi-scale approach that blurs the geometric model, rather than the image data, and show that this approach turns a global, highly nonquadratic optimization into a sequence of local, approximately quadratic problems that converge to the global minimum. The result is a deterministic, robust, and efficient optimization strategy applicable to a wide variety of Bayesian estimation problems in which geometric models of images are an important component.

We address the issue of how to make decision about the degree of smoothness demanded of a flexible contour used to model the boundary of a 2D object. We demonstrate the use of a Bayesian approach to set the strength of the smoothness prior for a tomographic reconstruction problem. The Akaike Information Criterion is used to determine whether to allow a kink in the contour.

Segmentation is an important step in the analysis of MR images (MRI). Considerable progress has been made in this area, and numerous reports on 3D segmentation, volume measurement and visualization have been published in recent years. The main purpose of our study is to investigate the power and use of fractal techniques in extraction of features from MR images of the human brain. These features which are supplemented by other features are used for segmentation, and ultimately for the extraction of a known pathology, in our case multiple- sclerosis (MS) lesions. We are particularly interested in the progress of the lesions and occurrence of new lesions which in a typical case are scattered within the image and are sometimes difficult to identify visually. We propose a technique for multi-channel segmentation of MR images using multiple feature vectors. The channels are proton density, T1-weighted and T2-weighted images containing multiple-sclerosis (MS) lesions at various stages of development. We first represent each image as a set of feature vectors which are estimated using fractal techniques, and supplemented by micro-texture features and features from the gray-level co-occurrence matrix (GLCM). These feature vectors are then used in a feature selection algorithm to reduce the dimension of the feature space. The next stage is segmentation and clustering. The selected feature vectors now form the input to the segmentation and clustering routines and are used as the initial clustering parameters. For this purpose, we have used the classical K-means as the initial clustering method. The clustered image is then passed into a probabilistic classifier to further classify and validate each region, taking into account the spatial properties of the image. Initially, segmentation results were obtained using the fractal dimension features alone. Subsequently, a combination of the fractal dimension features and the supplementary features mentioned above were also obtained and the results were compared with each other as well as with other standard techniques. A number of controlled experiments were carried out with the help of trained radiologists, which have indicated the validity of the preliminary results.

In this paper, we propose a methodology for evaluating medical image segmentation algorithms where the only information available is boundaries outlined by multiple expert observers. In this case, the results of the segmentation algorithm can be evaluated against the multiple observers' outlines. We have derived statistics to enable us to find whether the computer-generated boundaries agree with the observers' hand-outlined boundaries as much as the different observers agree with each other. We illustrate the use of this methodology by evaluating image segmentation algorithms on two different applications in ultrasound imaging. In the first application, we attempt to find the epicardial and endocardial boundaries from cardiac ultrasound images, and in the second application, our goal is to find the fetal skull boundaries from prenatal ultrasound images.

A true 3D split-and-merge segmentation technique is compared with a direct MRI measurement method for the estimation of ventricular CSF volume. The segmentation algorithm incorporates simulated annealing and boundary elimination and its performance is also compared with slice-based applications of 2D methods. The direct method is corrected for errors due to image non-uniformity. Using a set of volunteer studies the average differences between the two methods was found to be 36%. From this comparison a new source of measurement error in the direct method, caused by the inclusion of non-ventricular CSF signal, is identified and corrected by reference to the 3D segmentation. Adjusting the direct method in this way we have reduced the difference to 6%, suggesting that algorithmic approaches provide accurate ventricular volume estimation.

The texture-based segmentation of x-ray images of focal bone lesions using topological maps is introduced. Texture characteristics are described by image-point correlation of feature images to feature vectors. For the segmentation, the topological map is labeled using an improved labeling strategy. Results of the technique are demonstrated on original and synthetic x-ray images and quantified with the aid of quality measures. In addition, a classifier-specific contribution analysis is applied for assessing the feature space.

Accurate assessment of intrathoracic airway physiology requires sophisticated imaging and image segmentation of the three-dimensional airway tree structure. We have previously reported a rule-based method for three-dimensional airway tree segmentation from electron beam CT (EBCT) images. Here we report a new approach to airway tree segmentation in which fuzzy logic is used for image interpretation. In canine EBCT images, airways identified by the fuzzy logic method matched 276/337 observer-defined airways (81.9%) while the fuzzy method failed to detect the airways in the remaining 61 observer-determined locations (18.1%). By comparing the performance of the new fuzzy logic method and that of our former rule-based method, the fuzzy logic method significantly decreased the number of false airways (p less than 0.001).

Magnetic resonance images (MRI) of the brain are segmented into tumor and non-tumor to provide an objective means for measuring the response of a tumor to therapy. This paper focuses on automatic feature extraction to improve brain tumor segmentation using genetic algorithms (GAs). Using the discovered features, a significant increase in accuracy of tumor segmentation is seen. Moreover, this improvement is consistent over a tumor volume outside the training data.

Segmentation of volumetric medical imagery often benefits from the incorporation of a-priori knowledge of anatomical structure into the problem domain. Our previous work involved the use of active surfaces based on spherical mappings of rectangular grids to represent the anatomical structure of the object of interest. We have developed an improved method using active surfaces based on multiple geodesics which can dynamically link together to form object models. Deformable geodesic models are topologically closed structures, making them a better approximation of the types of objects found in volumetric imagery. In addition, geodesic surfaces exhibit regular sampling intervals which facilitate collision detection between different surfaces, allowing constraints among different surfaces to be included in the segmentation scheme. In this paper we present an overview of our current active surface segmentation system. Initial results of the system as applied to the problem of the segmentation of the major sub-parts of the brain are also presented. The results show favorable potential for the use of geodesic-based surfaces as a means for the segmentation of closed objects from volumetric medical imagery.

We propose a method for the segmentation of cerebral sulci, representing them by surfaces. This method is based on the computation of the differential characteristics of MRI data. The computation of curvature information, using the Lvv operator, allows one to differentiate sulcal and gyral regions, resulting in a global detection of the cortical scheme. The analytical description of a particular sulcus is obtained by initializing an active model on its trace upon the brain surface. The result is a surface representing the buried part of the sulcus. The 'snake-spline' model allows one to define an algorithm which is simpler and more robust than the classical snake. This method of segmentation yields good results for the 3D segmentation and visualization of cortical sulci.

A new methodology is proposed for the delineation of anatomical or physiological target regions utilizing a rich set of multimodality imagery. The proposed technique is unique in that it allows the integration of anatomical cues from transmission tomography (MRI/CT) with more fuzzy, physiological information derived from emission tomography (PET/SPECT). This approach allows for the delineation of regions of homogeneous tracer uptake constrained by the boundaries of anatomical structures, and vice versa. An extension of deformable model segmentation techniques is presented which integrates both region and edge information from registered scans in a competitive manner. The proposed technique has been implemented for two-dimensional deformable models. Results are presented for the cases of unimodal MRI in which region and edge information is integrated, and for a multimodal case employing a registered MRI and PET scan.

In multidimensional imaging, there are, and will continue to be, situations wherein automatic image segmentation methods fail and extensive user assistance in the process is needed. For such situations, we introduce a novel user-steered image boundary segmentation paradigm under two new methods, live-wire and live-lane. The methods are designed to reduce the time spent by the user in the segmentation process providing tight user control while the process is being executed. The strategy to reach this goal is to exploit the synergy between the superior abilities of human observers (compared to computer algorithms) in boundary recognition and of computer algorithms (compared to human observers) in boundary delineation. We describe evaluation studied to compare the utility of the new methods with that of manual tracing based on speed and repeatability of tracing and on data taken from a large on-going application. We conclude that the new methods are more repeatable and on the average two timers faster than manual tracing. Live-wire and live-lane operate slice-by-slice in their present form. Their 3D and 4D extensions, which we are currently developing, can further reduce the total segmentation time significantly.

We present a new multiscale contour fitting process which combines information about the image and the contour of the object at different levels of scale. The algorithm is based on energy minimizing deformable models but avoids some of the problems associated with these models. The segmentation algorithm starts by constructing a linear scale-space of an image through convolution of the original image with a Gaussian kernel at different levels of scale, where the scale corresponds to the standard deviation of the Gaussian kernel. At high levels of scale large scale features of the objects are preserved while small scale features, like object details as well as noise, are suppressed. In order to maximize the accuracy of the segmentation, the contour of the object of interest is then tracked in scale-space from coarse to fine scales. We propose a hybrid multi-temperature simulated annealing optimization to minimize the energy of the deformable model. At high levels of scale the SA optimization is started at high temperatures, enabling the SA optimization to find a global optimal solution. At lower levels of scale the SA optimization is started at lower temperatures (at the lowest level the temperature is close to 0). This enforces a more deterministic behavior of the SA optimization at lower scales and leads to an increasingly local optimization as high energy barriers cannot be crossed. The performance and robustness of the algorithm have been tested on spin-echo MR images of the cardiovascular system. The task was to segment the ascending and descending aorta in 15 datasets of different individuals in order to measure regional aortic compliance. The results show that the algorithm is able to provide more accurate segmentation results than the classic contour fitting process and is at the same time very robust to noise and initialization.

Texture analysis using convolution filters has been proposed as a method for the quantitation of extended objects in mammography. The work reported incorporates additional steps of image processing in an attempt to improve the performance of these filters. By deriving approximate outlines which are used to identify suspicious regions, the investigation illustrates the properties of one of the filters. After applying the method to a small pre-diagnosed database of stellate lesions and regular masses, the results show that the filter is able to outline the masses in all cases presented. The erroneous areas extracted are small for the initial part of the work which deals with 256 by 256 pixel image extracts, though slightly larger in some cases when the whole mammogram is considered. For each non-suspicious case studied, the sum of any false positive areas is statistically small when compared with the regions correctly outlined in the pre-diagnosed instances. By using an active contour model the outline match has been improved for the regular masses and also, but to a lesser degree, for the stellate lesions.

In this paper we present a technique for detecting and tracking the mitral leaflet boundaries in echocardiographic sequences of images covering a complete cardiac cycle. The developed method is based on the active contour framework. This approach requires initial placement of the contour in vicinity of a feature of interest (a leaflet boundary in this case). Contour shape and position are then modified by the energy minimization process. Active contour can be used for tracking by selecting the optimal solution from the previous frame as an initial position in the present one. Such approach fails when displacements of the feature between two consecutive frames are too large as is the case with mitral valve leaflets. We propose a technique for tracking leaflets and other fast moving objects that takes advantage of the velocity information obtained by optical flow estimation. Singh's multiresolution optical flow estimation technique is used since it can estimate large velocities. The algorithm has been tested on several sequences from transesophageal echocardiography. The advantage of utilizing velocity information proves to be important in frames where leaflet motion is large. The contour inertia and stiffness suppress sporadic inaccuracies in velocity estimates which results in a well-behaved contour.

The size and number of follicles present within an ovary may be used as an indicator of fertility in women. Ultrasound is the imaging modality of choice for obtaining information on the follicles as it is inexpensive and readily available. A method of segmenting the follicles and ovary and producing accurate 2D and 3D representation would be of great benefit to a large segment of the population. However, the nature of ultrasound images means that standard approaches to segmentation based on image gradients or detecting regions of homogeneous gray-level alone are inadequate. A semi-automatic method of segmentation which combined a texture based classification for initial segmentation with deformable models to provide descriptions of individual objects is extended by imposing geometric constraints on the relationships between the individual objects present within an image. Since we are interested in segmenting the individual objects over a 3D spatial stack we use the results from one image in the sequence as the initial estimates for the next image. This reduces the need for operator intervention and provides representations of individual objects through the whole sequence. These representations can then be used for accurate measurement of area/volume and for three-dimensional visualization of the relationships between the individual follicles and the enclosing ovary.

All retrospective image registration methods have attached to them some intrinsic estimate of registration error. However, this estimate of accuracy may not always be a good indicator of the distance between actual and estimated positions of targets within the cranial cavity. This paper describes a project whose principal goal is to use a prospective method based on fiducial markers as a 'gold standard' to perform an objective, blinded evaluation of the accuracy of several retrospective image-to-image registration techniques. Image volumes of three modalities -- CT, MR, and PET -- were taken of patients undergoing neurosurgery at Vanderbilt University Medical Center. These volumes had all traces of the fiducial markers removed, and were provided to project collaborators outside Vanderbilt, who then performed retrospective registrations on the volumes, calculating transformations from CT to MR and/or from PET to MR, and communicated their transformations to Vanderbilt where the accuracy of each registration was evaluated. In this evaluation the accuracy is measured at multiple 'regions of interest,' i.e. areas in the brain which would commonly be areas of neurological interest. A region is defined in the MR image and its centroid C is determined. Then the prospective registration is used to obtain the corresponding point C' in CT or PET. To this point the retrospective registration is then applied, producing C' in MR. Statistics are gathered on the target registration error (TRE), which is the disparity between the original point C and its corresponding point C'. A second goal of the project is to evaluate the importance of correcting geometrical distortion in MR images, by comparing the retrospective TRE in the rectified images, i.e., those which have had the distortion correction applied, with that of the same images before rectification. This paper presents preliminary results of this study along with a brief description of each registration technique and an estimate of both preparation and execution time needed to perform the registration .

In this paper we quantitatively evaluate the amount of residual registration error for different settings of an internal parameter of our surface-based multimodality registration system. The investigated parameter controls the accuracy in representing one of the surfaces to be registered. Computational complexity of our registration system is also directly related to the surface representation accuracy for a given data set. We measured registration error using a previously developed reference data set, which accurately determines registration error at many locations throughout the brain. We found a smaller registration error when the registered surface was represented more accurately. The difference in error was not as large as anticipated, suggesting that less accurate representations can be used by our system.

Image registration is a valuable technique for medical diagnosis and treatment. It allows physicians to combine information from multiple images by aligning them into the same coordinate space. Surface-based methods register images by aligning corresponding surfaces of one or more anatomical structures such as the surface of the brain or skull. In this paper we examine a novel implementation of this approach. We use the iterative closest point algorithm to iteratively search for the transformation that minimizes the distance between surface points in one image and a surface model in the other image. In each iteration we use the closest point transform to find corresponding points that are closest with respect to the transformation estimated in the previous iteration. Results from several experiments are presented to demonstrate the efficacy of this approach.

Correlation provides rapid, accurate, objective image registration, but is limited to applications where the data sets are similar. Our goal was to develop preliminary image processing techniques which will allow correlative registration of (dissimilar) brain images obtained using different imaging modalities. A simulation showed that correlative registration is effective for trend-similar images, leading to the hypothesis that preprocessing algorithms need only impose trend similarity to permit successful cross-modality correlative registration. This hypothesis was tested using MR and CT image which were acquired with external fiducial markers to test the registration accuracy. After analyzing the salient features of each image set, trend similarity was imposed using simple masking and gray scale adjustment techniques. A correlative registration program was then applied to the trend-similar image sets, and the resulting rigid body transformation parameters were applied to the original images. The registered image sets matched the fiducial marker locations to within one pixel, accuracy is similar to that obtained by matching user-identified anatomical points, leading to the conclusion that simple pre-processing techniques can allow correlative registration to be used for aligning images from multiple imaging modalities.

Noise exists in every practical signal, but it is seldom taken into consideration when one studies or performs interpolation. In this paper, we investigate the effect of noise on the performance of an interpolating function. Using this study, we propose a distortion measure that can be used as a basis for selecting an interpolating function for a practical signal. Attention is given to both uncorrelated noise and CT (computed tomography) noise. We have developed an algorithm to estimate this measure from representative images, and use both simulated and real CT images to show that selection of the interpolating function based on this distortion measure leads to a more accurate and efficient interpolation.

Spatial registration is a major problem arising whenever several images of similar contents are to be compared. Considering translations only, two-dimensional cepstral techniques have been proven to be exact and robust against noise or intensity variations. Furthermore, the cepstral filtering is numerically more efficient than most common approaches to image registration based on cross-correlation or template matching. In a previous paper, we proposed a two- dimensional cepstrum based matching technique accessing rotations and translations. The logarithmic polar mapping of the power spectra of both images to be registered is used for the decoupling of rotations and translations (similar to the Fourier-Mellin transform). Rotations are detected first matching the mapped spectra by two-dimensional cepstrum analysis. After rotating back one image, the relative shift is determined using the same cepstrum technique. In clinical practice, the rotation detection step was discovered as the weakness of this registration technique. Based on 855 pairs of dental radiographs acquired in known positions, three different approaches of matching the mapped spectra are compared: the cepstrum technique, the cross-correlation, and the entropy of the one-dimensional histogram distribution function of the substraction image of the mapped spectra. The combination of the log-polar mapped power spectra of both x rays with the entropy-measure allows the best detection of rotations. The union with common cepstrum methods correcting translations results in a robust rotation- extended cepstrum technique.

Volume morphing is a technique used for generating smooth 3D image transformations and deformations. In this paper, several algorithms are developed for morphing transformations that create new forms and simulate shape deformation and growth in biomedical applications. Three-dimensional biological point landmarks and their movements are defined to guide the morphing transformation using a scattered data interpolation algorithm. Two types of morphing algorithms, the Shepard-based interpolation and the radial basis function approach, are investigated in detail and applied for volume morphing in growth predictions of children and shape transformations between species.

The Gibbs priors with potential equal to the membrane deflection and thin plate bending energies are explored in the Bayesian approach to image matching. Their smoothness properties are qualitatively demonstrated in a matching task. The priors are further evaluated by comparing their effect on the atlas-based localization of several subcortical structures in MRI data. Results of the localization study indicate that the implementation based on the membrane prior assumed over a fine mesh outperforms, both in speed and accuracy of the anatomic labeling, a plate-based approach that uses a comparable number of unknowns.

Spiral CT-angiography (CTA) is a new, minimally invasive technique for vascular imaging. A key task in CTA is to suppress bone and other dense tissues, such as calcification, in order to visualize the vessels of interest. Common practice in CTA employs automatic or semiautomatic bone editing techniques. However, editing is not possible where the bone structures are complex, and particularly when they are close to vessels. In addition, it is quite unlikely that any automatic editing technique would succeed in removing calcification. An alternative approach is to extend conventional digital subtraction techniques (DSA) from two dimensions to three dimensions. The potential advantages of digital subtraction CTA are; (1) fully automatic bone removal, (2) elimination of calcification, and (3) reducing the concentration of contrast medium. This paper presents a 3D flexible registration technique that brings a pre-contrast (native) volume to register with a post-contrast (contrast) volume prior to subtraction. Two volumes are treated to be composed of blocks; each block in one data set is translated, rotated, and scaled in three dimensions to match the corresponding block of the other volume. In essence, flexible matching is realized by rigid body motion and scaling of a large number of blocks. The proposed algorithm has been successfully applied on a number of clinical studies.

We present an automatic and unsupervised method for non-rigid registration of 3D magnetic resonance (MR) images with the Stockholm Computerized Brain Atlas (CBA). This method can be used in the context of multimodal medical image registration, fusion and automatic brain segmentation. In these applications anatomical images (MR) are coregistered with low spatial resolution functional imaging modalities (PET and SPECT) and fused with the neurological database of the CBA. The proposed matching method is based on the minimization of a 3D Chamfer distance function between the surface of the brain extracted from the MR image and the CBA brain surface. The surface-to-surface distance function is efficiently calculated by using a precomputed point-to-surface Euclidean distance map. The non-rigid inter-patient transformation of the CBA is modeled by a generalized 3D second order transformation. This transformation is easily differentiable and, as a consequence, fast and efficient minimization methods can be used. First, a quasi-rigid, first order transformation is computed. Then, the matching is improved by introducing the second order coefficients into the transformation. After this global matching, a local adaptation of the CBA is performed by a morphing method. The combination of a second order global transformation with a 3D local morphing allows the user to obtain a registration accuracy of one pixel, i.e. a mean distance between the surface of the brain in the MR image and the CBA of one pixel, which is significantly better than what can be expected from a human operator.

An adaptive Wiener-filter image restoration approach is being described. It is exemplarily applied to intraoral radiographs generated by two x-ray-sensitive CCDs, one having an additional scintillator screen. The devices' MTFs were determined by means of a slightly angulated slit according to DIN 6867/2. Noise power spectra at various exposures were measured and decomposed into quantum and detector noise with respect to the signal and noise model. Both noise power levels were proportional to the incident exposure. Wiener filters for high, low and medium exposures were designed. Weight tables for superpositions of these filters were calculated by minimizing the mean square errors between the superposition MTFs and Wiener filter MTFs for arbitrary exposures. The pixelwise weights depending on the local mean were calculated after a preceding lowpass filtering. The final restored images were composed from spatially invariant filtered images generated by 2D-IIR filtering. The results indicate that the presented adaptive restoration method is a promising approach with regard to both image quality and processing speed demands.

Two techniques for pincushion correction are evaluated based on their effect on calculation of the image geometry and 3D positions of object points. Images of a uniform wire mesh and a calibration phantom containing lead beads in its surface were acquired on the image intensifier TV systems in our catheterization labs. The radial mapping functions relating points in the original images and in the corrected images were determined using the mesh image. The undistorted mesh model was also used to determine and correct the distortions locally, i.e., for each square region between the mesh points. Thus, two corrected images were obtained. Images of the calibration phantom before and after correction were analyzed to determine the 3D position of the lead beads and the imaging geometry, using a calibration algorithm and the enhanced Metz-Fencil technique. In comparing the 3D positions calculated from the radially corrected and locally corrected images, the calculated 3D positions using the calibration technique vary by less than 0.6 mm in the x and y direction and less than 5.0 mm in the z direction. The uncorrected data yields differences of over 1 cm in the z direction. The 3D positions calculated using the enhanced Metz-Fencil technique appear to be more accurate when pincushion correction is applied.

Biomedical image data, such as obtained from CT or MR imaging modalities, tend to occupy a large amount of storage space. At the cost of losing or distorting salient features, lossy compression techniques can be used to reduce significantly the amount of data storage space. The main aim of this article is to present a novel method for optimal enhancement and restoration of images recovered from data stored using lossy compression techniques. A statistical model of the deformations undergone by the salient features within the original image, when it is stored using lossy compression techniques and then recovered, is generated. This statistical model generates a discrete lattice space. The algorithm presented here designs a set of filters over the statistical lattice space. Due to the statistical nature of the lattice the designed filters are optimal and the best possible recovery of the salient features in the original image is achieved. Results comparing the performance of the presented method to that achieved by median filters are presented. Robustness of the algorithm is tested by applying filters generated using a set of images of one subject, to images of different subjects and images stored using different compression ratios.

This paper presents an image processing approach for correcting streaking artifacts in computed tomography (CT) images and compares it to physics based and composite approaches. In CT, there are two major sources for streaking artifacts: beam hardening and nonlinear partial volume averaging. The physics based approach uses a physical model for the beam hardening. Inaccuracy of the model and nonlinear partial volume averaging result in an incomplete correction of the artifacts by this approach. The proposed image processing approach identifies and corrects the artifacts using their frequency characteristics. It corrects the artifacts regardless of their source. The composite approach also utilizes the physical model for the beam hardening. Although correcting some of the artifacts, it is computationally more intense than the image processing approach. We illustrate the methods and compare their performance using a computer simulation and CT images of a phantom and a human brain. We show that the streaking artifacts resulting from nonlinear partial volume averaging which are not corrected by the physics based method are corrected by the image processing approach.

Due to coherent imaging techniques employed in medical imaging such as ultrasound imaging, signal dependent speckle noise appears as granular noise that is difficult to remove without blurring the image details. Considerable efforts have been spent to design systems with improved resolution and reduced speckle noise. This study investigates the effect of recently developed nonlinear and linear filtering techniques in restoring medical images corrupted with speckle noise. Preliminary results demonstrate that a class of nonlinear filters based on mathematical morphology outperforms linear wavelet filters and a simple nonlinear median filter in removing speckle noise.

Accurate and reproducible geometric measurement of lower extremity residua is required for custom prosthetic socket design. We compared spiral x-ray computed tomography (SXCT) and 3D optical surface scanning (OSS) with caliper measurements and evaluated the precision and accuracy of each system. Spiral volumetric CT scanned surface and subsurface information was used to make external and internal measurements, and finite element models (FEMs). SXCT and OSS were used to measure lower limb residuum geometry of 13 below knee (BK) adult amputees. Six markers were placed on each subject's BK residuum and corresponding plaster casts and distance measurements were taken to determine precision and accuracy for each system. Solid models were created from spiral CT scan data sets with the prosthesis in situ under different loads using p-version finite element analysis (FEA). Tissue properties of the residuum were estimated iteratively and compared with values taken from the biomechanics literature. The OSS and SXCT measurements were precise within 1% in vivo and 0.5% on plaster casts, and accuracy was within 3.5% in vivo and 1% on plaster casts compared with caliper measures. Three-dimensional optical surface and SXCT imaging systems are feasible for capturing the comprehensive 3D surface geometry of BK residua, and provide distance measurements statistically equivalent to calipers. In addition, SXCT can readily distinguish internal soft tissue and bony structure of the residuum. FEM can be applied to determine tissue material properties interactively using inverse methods.

A high-resolution 3D radiological image provides a virtual copy of the anatomy that can be used as an input to a computer-based image-analysis system. In particular, useful information can be obtained by interactively navigating through a 3D radiological image in a manner similar to an endoscopic examination. Traditional endoscopy, though, only provides views in restricted regions (e.g., inside hollow passages) and only offers limited exploration and quantitation. We present a computer-based endoscopic system, QUICKSEE, that heads toward overcoming these difficulties. QUICKSEE enables noninvasive unrestricted exploration of the virtual anatomy and can be useful as an anatomical quantitation tool. The system is particularly suited for analyzing complex 3D branching structures such as the pulmonary airways and coronary arteries. The paper describes the current form of QUICKSEE and provides examples of the system's use for 3D pulmonary analysis and 3D microCT imaging.

Quantifying changes in the number and extent of lesions in MR images has been used to indicate disease activity in multiple sclerosis. However, quantification is often long and arduous. This has prompted research into new quantification and image processing algorithms to speed up and simplify lesion quantification. Nevertheless, many algorithms rely upon an experienced operator, and are thus subject to inter- and intra-operator variability. We present a new technique to measure operator variability which is largely independent of the lesions selected for analysis, and is expressed in the measurement units. This new measure allows researchers to predict and compare absolute uncertainty for operators using new algorithms. We used this technique to examine the impact on operator performance of computer assisted lesion quantification and anisotropic filtering of patient exams to reduce image noise. In both cases repeated measurements of lesions in MR exams were performed and analyzed. Results indicate that assisted quantification reduced inter-operator variability by half (from 0.34 cm³ to 0.17 cm³) and reduced intra-operator variability by one-third (from 0.23 cm³ to 0.15 cm³). Anisotropic filtering reduced inter- and intra-operator variability by one-fifth (from 0.34 cm³ to 0.27 cm³, and from 0.23 cm³ to 0.19 cm³, respectively). These results may have practical implications for clinical trials which rely on quantitative measurements to assess therapeutic effect.

Conventional nuclear medicine images are spoiled by photon attenuation and scattering. Decreased contrast, blurred object edges and erroneous quantification are the most obvious consequences. Both processes are intimately linked so that a proper correction can hardly be achieved. We have investigated the usefulness of a neural network based approach (ANN) to compensate for these damages. Numerical Monte-Carlo simulations and physical phantoms acquisitions of homogeneous sources of various forms and volumes in a diffusing medium were used to examine these capacities. Using the energy spectrum of incident photons for every pixel of each image and two diametrically opposed views of the radioactive objects as sources of information, a multilayer neural network with backpropagation as learning tool, we were able to get a proper restitution of images so that it seems now possible to run meaningful quantification.

Grayscale inhomogeneities in magnetic resonance (MR) images cause significant problems in automated quantitative image analysis. Removal of such inhomogeneities is a difficult task, but it has been investigated by a number of different authors recently. The most common methods used involve some type of homomorphic filtering to create a smoothed version of the original image, which is then used as an estimate of the bias field to be removed from the image. Many investigators have implemented variations of this technique and have demonstrated their usefulness for a wide range of applications, but no investigator has yet attempted a systematic, quantitative study to describe the effects these algorithms have on images. This study introduces a quantitative paradigm for evaluating inhomogeneity correction algorithms by their performance on a constructed simulation image with different bias fields applied. We find that mean filter algorithms are more successful than median filter algorithms, and that larger kernel sizes than what are currently reported in the literature offer significant improvements in post-correction image quality.

The aim of this work is the development of a non-invasive technique for efficient and accurate volume quantization of the cerebellum of mice. This enables an in-vivo study on the development of the cerebellum in order to define possible alterations in cerebellum volume of transgenic mice. We concentrate on a semi-automatic segmentation procedure to extract the cerebellum from 3D magnetic resonance data. The proposed technique uses a 3D variant of Vincent and Soille's immersion based watershed algorithm which is applied to the gradient magnitude of the MR data. The algorithm results in a partitioning of the data in volume primitives. The known drawback of the watershed algorithm, over-segmentation, is strongly reduced by a priori application of an adaptive anisotropic diffusion filter on the gradient magnitude data. In addition, over-segmentation is a posteriori contingently reduced by properly merging volume primitives, based on the minimum description length principle. The outcome of the preceding image processing step is presented to the user for manual segmentation. The first slice which contains the object of interest is quickly segmented by the user through selection of basic image regions. In the sequel, the subsequent slices are automatically segmented. The segmentation results are contingently manually corrected. The technique is tested on phantom objects, where segmentation errors less than 2% were observed. Three-dimensional reconstructions of the segmented data are shown for the mouse cerebellum and the mouse brains in toto.

The medial axis analysis of an object can be used to effectively guide and optimize radiosurgery treatment planning. In this paper, a fast Euclidean medial axis transformation in three dimensions based on dynamic grassfire simulation and ridge extraction is presented. A ridge occurs when fire fronts collapse during grassfire propagation. Iso-contours (2D) or iso- surfaces (3D) can be obtained from dynamic grassfire transforms. They are locally smooth everywhere except at ridge locations. Ridges are detected by measuring local curvature at each point. This process is invariant under spatial translations and rotations. The algorithm yields the true Euclidean skeleton of the objects and is several orders of magnitude faster than other thinning methods. In radiosurgery treatment planning, optimal shots are only placed on the medial axis of the 3D target, which reduces optimization time and complexity. An example of a treatment planning process is presented and the relationship between skeleton disks and the dose distributions which they predict are discussed.

The construction of a multiscale (scale space) representation requires the smoothing of the given image to generate a set of corresponding images at other coarser scales, and the extraction of features at these scales. Various methods of smoothing, combined with various feature extractors, will result in drastically different scale space representations. The particular application and desired criteria determine the choice of smoothing and feature extractor. In this paper, we first put foreword a framework for the scale space representation. Then we focus on the Gaussian and morphological scale space for planar shape analysis of computed radiographic (CR) images. The discussion of the various scale space methods is organized into three categories -- boundary approach, region approach, and hybrid approach. Properties, limitations, performance, and applications of these scale space methods are discussed. Extensive experiments on CR images of various exam types were conducted and the results are evaluated.

In dental implantology more than one hundred enossal implant systems are in use. Once embedded, the dental x-ray examination is the most important tool for determining implants' producer, name, and type. In this paper, we present a system for automatic detection and identification of dental fixtures in intraoral x rays (IDEFIX) combining common direct digital image acquisition techniques with specially designed image analysis. IDEFIX can process any digital radiograph (e.g. RVG, Sens-A-Ray, Schick, Sidexis, Digora) as well as digitized dental films. A reference database has been generated by precise measurement on the implant systems used so far (eight implants) including parameters like length, diameter, and cross section area. After binarization of the current digital x-ray image, a parameter set is extracted from each detected object applying mathematical morphology. All objects are classified using a simplified nearest neighbor method and the Euclidean distance metric. If the distance of the objects' parameter set to one of the reference sets is below a given threshold, name and type of the identified dental fixture are displayed on the screen. Otherwise, the actual object will be rejected as a no-implant. IDEFIX has been evaluated by processing various in-vitro acquired radiographs. Different implants were classified captured with identical conditions as well as acquired varying the angulation of the x-ray tube. It is shown that misangulations up to twenty degrees are tolerable preserving correct identification. Other image structures like teeth or fillings result in large distances to all reference parameter sets and, therefore, they are reliably recognized as non-implants.

In this paper we present a hierarchical multiscale shape description tool based on active contour models which enables data-driven quantitative and qualitative shape studies of MR brain images at multiple scales. At large scales, global shape properties are extracted from the image, while smaller scale features are suppressed. At lower scales, the detailed shape characteristics become more prominent. Extracting a shape at different levels of scale yields a hierarchical multiscale shape stack. This shape stack can be used to localize and characterize shape changes like deformations and abnormalities at different levels of scale. The shape description is performed as a set of implicit segmentation steps at multiple scales yielding descriptions of an object at various levels of detail. Implicit segmentation is carried out using the well-known model of active contours. Starting from an initial active contour, several implicit optimization processes with differently regularized energy functions are performed, where the energy functions are represented as functions of scale. The presented algorithm for shape focusing and description based on active contour models shows promising results on extracting and characterizing complex shapes in MR brain images at a large set of scales.

Active contour models have become popular recently especially in medical image segmentation research and in the modeling of deformable structures. There are a number of theoretical, computational, and operational advantages to treating the boundary as a digital entity and defining deformations digitally on the digital representation. Nevertheless, this does not seem to have been pursued in the literature. A significant body of knowledge exists in the literature on the theory and algorithms related to nD, oriented, closed, connected, digital surfaces. In this paper we examine how this knowledge can be utilized to extend active contour models to nD digital surfaces. We present an algorithm for deforming nD, oriented, closed, connected digital surfaces into nD, oriented, closed, connected digital surfaces. The algorithm uses a data structure which we call a random insertion/deletion list. It contains spatial elements (spels) that constitute the immediate interior (II) of the surface as well as information that establishes the relationship of the spels to the surface elements (surfels). (In our digital surface model, the surfels are one lower dimensional than the spels. For example, in 3D, the spel is a voxel and a surfel is a square representing the voxel face). All deformations are realized essentially by adding/deleting spels to/from the data structure. Our results indicate that it is possible to model surfaces digitally and that given an nD, oriented, closed, connected, digital surface, we can deform this surface using local operations only to produce an nD, oriented, closed, digital surface. We also describe an algorithm and data structures to implement this model efficiently. Furthermore, if we allow more global tests, we can also insure that connectedness of the surface is insured as well. We conclude that the extensively studied digital surface models are appropriate for representing deformable objects also. Because of the finite (and small number of) local configurations of surfels and spels, these models offer efficient deformation computation as has been repeatedly demonstrated in the past for a variety of visualization and analysis operations. We demonstrate that we can deform the surface in such ways that maintain orientation and closure using only local checks and we can also insure connectedness at the expense of global checks.

A newly developed magnetic resonance imaging technique can directly visualize propagating acoustic strain waves in tissue-like materials. By estimating the local wavelength of the acoustic wave pattern, quantitative values of shear modulus can be calculated and images generated that depict tissue elasticity or stiffness. Since tumors are significantly stiffer than normal tissue (the basis of their detection by palpation), this technique may have potential for 'palpation by imaging,' with possible application to the detection of tumors in breast, liver, kidney, and prostate. We describe the local wavelength estimation algorithm, study its properties, and show a variety of sample results.

The application of inverse scattering methods to diagnostic ultrasound echo signals has provided us with detailed information about renal microstructure and function. In particular, the average scatterer size has been used to follow changes in microvascular perfusion that occur early in many renal disease processes. This paper shows that by introducing prior knowledge of the tissue state into the process, uncertainty in the spectral estimate is reduced for low SNR situations, and the contrast and range-resolution in scatterer size images can be improved without increasing the noise. Prior information used in the estimation technique is obtained from the histology of biological tissue. Maximum a posteriori and constrained least squares estimators are designed to obtain images for different levels of noise and for different gate-durations. Prior knowledge about the noise properties and the nature of the echo spectrum is used to obtain the order of an autoregressive model for estimating the power spectral density.

In this work we encode the shape complexity of a search object using 3D Fourier descriptors (FDs) and allow genetic algorithms (GAs) to optimize the object's shape and position. Using magnetic resonance image (MRI) data, we perform an approximate segmentation on one lateral ventricle in the brain and use the FDs from this as seeding values for the GAs to search for the left and right lateral ventricles in subsequent 3D data sets. We show that the method is capable of coping with normal biological variation. We compare a GA-guided segmentation with an interactive region growing method and find an agreement of not less than 80 plus or minus 6% in voxel classification with a corresponding average edge placement error of 2.2 plus or minus 0.4 mm. Finally we examine how the optimization can be speeded up by a distributed parallel implementation.

In this paper we present 3D registration techniques for detecting breast cancer lesions based on differences between the high spatial resolution 'pre' and 'post' contrast administration MR breast scans. We also present registration techniques for detecting lesions based on the time course of enhancement after administration of the contrast agent based on changes in the low spatial resolution, high temporal resolution, dynamic MR data taken during the absorption of the contrast agent. Alignment of the 'pre' and 'post' MR data and of the dynamic MR images is done by a direct optimization technique to estimate a global affine deformation model using a coarse-to-fine control strategy over a 3D pyramid. This global model is followed by a one iteration flow estimation to account for any local non-rigidity. We present results of the registration process and visualization of the difference volumes obtained after alignment.

An approach for speckle reduction and feature enhancement under a framework of multiscale wavelet analysis is presented. The advantages of both soft thresholding and hard thresholding wavelet shrinkage techniques are utilized to eliminate noise and preserve the sharpness of salient features. We integrate a method of wavelet shrinkage with nonlinear processing to enhance contrast within structures and along object boundaries. Feature restoration and enhancement are accomplished by modifying the gain of a signal's variational energy. In this study, we focus on multiplicative noise, such as speckle noise. We show that this algorithm is capable of enhancing features of interest, such as endocardial and epicardial boundaries in 2-D short-axis echocardiograms while at the same time reducing speckle. Speckle is modeled as multiplicative noise, and approximated by additive stationary Gaussian white noise on a logarithmic scale. In our algorithm, shrinkage of wavelet coefficients via soft thresholding within finer scales is carried out on coefficients of logarithmically transformed echocardiograms. Nonlinear feature emphasis combined with hard thresholding within middle levels of transform space are performed subsequently on wavelet coefficients. The first operation reduces speckle while the second step accomplishes structure and object boundary enhancement. Preliminary results suggest that the algorithm can remove speckle noise and enhance contrast along boundaries of importance to cardiologists. We have applied this algorithm to echocardiograms of varying quality, and present both experimental analysis and sample results.

Focal bone lesions and bone tumors are of special interest in radiology because of their rare appearance (only one percent of all tumor diseases). This motivates a computer-assisted diagnosis recognizing bone tumors. Our image analysis extracts the radiomorphologic features in x rays using a texture-based approach. Diagnosing x rays, the radiologist examines regions of different size in x rays to gain both local and global impressions of the morphologic structure. In order to analyze the x ray in different resolutions, a multiresolution approach based on the wavelet transform is applied to the radiographs. To measure the informational content of the wavelet coefficients for the individual morphologic structures, we calculated a normalized summation of the absolute wavelet coefficients within a local N by N window and called this feature the local energy. We proved in different tests this feature and the parameter for calculating the wavelet transform for a correct classification of the medical structures, applying a topologic map from Kohonen. It is shown that the wavelet transform is well suited for the feature extraction of textures.

A hierarchical Markov random field (MRF) modeling approach is presented for the classification of textures in selected regions of interest (ROIs) of chest radiographs. The procedure integrates possible texture classes and their spatial definition with other components present in an image such as noise and background trend. Classification is performed as a maximum a-posteriori (MAP) estimation of texture class and involves an iterative Gibbs- sampling technique. Two cases are studied: classification of lung parenchyma versus bone and classification of normal lung parenchyma versus miliary tuberculosis (MTB). Accurate classification was obtained for all examined cases showing the potential of the proposed modeling approach for texture analysis of radiographic images.

Typical digital subtraction angiography (DSA) acquisition rates are often inadequate for visualizing and analyzing fast-moving flow patterns. Therefore, an interpolation method that captures the angiographic flow pattern was developed. The temporal change of gray value in each pixel along a blood vessel records the flow movement at that location. Thus, temporal interpolation was performed on a pixel-by-pixel basis. To generate each interpolated image, a polynomial interpolation was applied to six sequential images. To validate the interpolation technique, a flow phantom was imaged with a high acquisition frame rate, and interpolation was done in a lower frame rate and compared to the acquired data. The interpolated images were also compared to results from linear interpolation and cubic spline interpolation. Clinical utility was illustrated on DSA images of cerebral vasculature with aneurysms. Image sequences of 60 frame/s were generated from DSA images acquired at 7.5 frame/s. The results show improved flow pattern visualization, especially flow head locations in blood vessels. This interpolation method has also been applied to dynamic 3D reconstruction from biplane DSA projections. In this application, the method was used to offset temporal discrepancies between biplane projection pairs and contrast injections, making dynamic 3D reconstruction possible.

This paper presents a method of enhancing linear and curvilinear image features, such as those corresponding to tissue discontinuities in medical ultrasound. The method is an extension of a template based technique for line enhancement which produces a test statistic at each point by projecting the pixels near that point onto a line segment, varying the orientation of the segment to maximize the projected value, and retaining the projected value as the test statistic. In the past, we have not made use of information about which angle produced the maximum value at each point. In this paper, we compute a histogram of the angles near each point to gain an indication of the direction of larger scale linear features lying nearby. Mathematically, we wish to estimate a set of prior probabilities for the orientation of line segments that pass through each point. The priors can then be used to improve the power of the Bayesian line detection procedure. In addition, they can also be used to improve the visual quality of the image produced by plotting the test statistics on an image raster. We have found that such an image is revealing because it shows more sharply the edges of the linear components, making them more clearly visible and their fringes more distinguishable from the background. With the incorporation of prior information, the processed image shows a further improvement in visual and machine detectability of linear components, due to increased difference in gray level between points lying on edges and those lying away. This technique has potential to significantly improve the machine detectability of tissue discontinuities in medical ultrasound, as well as linear features in other forms of computed imaging.

Previously, we have shown the effectiveness of using Bayesian image estimation (BIE) to reduce scatter and increase the contrast to noise ratio (CNR) in digital chest radiography. Here, we investigate the use of BIE to reduce scatter and increase CNR on digital mammographic data. Calibrated photostimulable phosphor digital images were obtained of the American College of Radiologists (ACR) mammographic phantom at several different exposures. An iterative Bayesian estimation algorithm was used to process this data. Residual scatter fractions (RSF) and CNR were computed. Resolution was visually inspected. These results were compared to those of a mammogram acquired at standard clinical imaging parameters using an anti-scatter grid for scatter reduction. On average, at all exposure levels. BIE reduced scatter fractions from 57% to 6%, while a grid only reduced RSF to 31%. At similar exposure levels, BIE processing improved CNR to 21.6, while a grid produced images with a CNR of 15.8. At an exposure level of 37% less than the standard exposure, BIE improved CNR to 18.9. A visual assessment of resolution using the objects in the phantom showed no reduction of resolution. In some images, phantom masses appeared more readily apparent. BIE processing of mammographic data can reduce scatter and increase image CNR. This type of image processing may potentially allow for decreased radiation dose to the patient with no loss of image quality. BIE as a method for scatter compensation in mammography is very promising. This preliminary work shows improvement in CNR to values greater than that of a standard grid.

We have developed a visualization environment, CEPH, for the analysis of plain film, biothogonally registered, head x-rays, i.e., cephalograms. Most importantly this program facilitates the collection of 3D landmark data from biorthogonal cephalogram pairs, i.e., correlated frontal and lateral images. To this end we have implemented tools for: contrast enhancement, image compression, morphological feature extraction, and grayscale shape recognition for automatic landmark detection. These data are useful for clinical craniofacial: (1) diagnosis, (2) treatment planning, (3) computed-assisted surgery, and (4) post-procedure follow-up. We are currently using the CEPH visualization environment to produce average landmark sets from the nearly 2000 'Bolton Standards' images. These images were collected for the Bolton-Brush Growth Study at Case Western Reserve University. Two dimensional landmark data captured from these images are universally recognized and used as diagnostic and treatment 'norms.' We are planning to make new three dimensional Bolton Standards data available this year. To this end we have scanned all of the Bolton Standards cephalograms at 2400 by 2400 resolution with 12 bits of grayscale information on a TDI scanner.

The study aim is to develop an artificial neural network (ANN) for computer-aided diagnosis of mammography. Using 9 mammographic image features and patient age, the ANN predicted whether breast lesions were benign, invasive malignant, or noninvasive malignant. Given only 97 malignant patients, the 3-layer backpropagation ANN successfully predicted the invasiveness of those breast cancers, performing with Az of 0.88 plus or minus 0.03. To determine more generalized clinical performance, a different ANN was developed using 266 consecutive patients (97 malignant, 169 benign). This ANN predicted whether those patients were benign or noninvasive malignant vs. invasive malignant with Az of 0.86 plus or minus 0.03. This study is unique because it is the first to predict the invasiveness of breast cancers using mammographic features and age. This knowledge, which was previously available only through surgical biopsy, may assist in the planning of surgical procedures for patients with breast lesions, and may help reduce the cost and morbidity associated with unnecessary surgical biopsies.

Microcalcifications are important cues used by radiologists for early detection in breast cancer. Individually, microcalcifications are difficult to detect, and often contextual information (e.g. clustering, location relative to ducts) can be exploited to aid in their detection. We have developed an algorithm for constructing a hierarchical pyramid/neural network (HPNN) architecture to automatically learn context information for detection. To test the HPNN we first examined if the hierarchical architecture improves detection of individual microcalcifications and if context is in fact extracted by the network hierarchy. We compared the performance of our hierarchical architecture versus a single neural network receiving input from all resolutions of a feature pyramid. Receiver operator characteristic (ROC) analysis shows that the hierarchical architecture reduces false positives by a factor of two. We examined hidden units at various levels of the processing hierarchy and found what appears to be representations of ductal location. We next investigated the utility of the HPNN if integrated as part of a complete computer-aided diagnosis (CAD) system for microcalcification detection, such as that being developed at the University of Chicago. Using ROC analysis, we tested the HPNN's ability to eliminate false positive regions of interest generated by the computer, comparing its performance to the neural network currently used in the Chicago system. The HPNN achieves an area under the ROC curve of Az equal to .94 and a false positive fraction of FPF equal to .21 at TPF equals 1.0. This is in comparison to the results reported for the Chicago network; Az equal to .91, FPF equal to .43 at TPF equal to 1.0. These differences are statistically significant. We conclude that the HPNN algorithm is able to utilize contextual information for improving microcalcifications detection and potentially reduce the false positive rates in CAD systems.

Automatic detection of blood vessel shapes locating in the skin has a great diagnostic importance. In this work, an evolutionary approach operating on morphological operator and operation structures is proposed for the determination of the shape and network of blood vessels located in upper skin layers. A population of individuals comprising morphological structures is generated. A two-dimensional queue like data representation of individuals is applied in order to provide an appropriate representation of the connectivity constraints originated in the two dimensional nature of the structuring elements. Two-dimensional crossover and mutation type manipulation operations are carried out on selected elements of the population. Unlike the usual techniques, in our approach no constraints are used for background and smoothness as no matched filter or linear operator is applied. Also no a priori knowledge of the vessel shape is necessary due to the evolutionary method. Unlike the usual imaging techniques, that mainly use angiograms as input, in this work infrared filtered images taken by CCD camera are applied to investigate the blood vessels of broad skin areas. The method is implemented parallel on a lattice network of transputers resulting in a significantly decreased processing time compared to the usual techniques.

The number of nuclei on a pathology image assists pathologists in consistent diagnosis of breast cancer. Currently, most pathologists make a diagnosis based on a rough estimation of the number of nuclei on pathology images. Because of the rough estimation, the diagnosis is not objective. To assist pathologists to make a consistent, objective and fast diagnosis, it is necessary to develop a computer system to automatically recognize and count several kinds of nuclei. We have developed an algorithm for the automatic segmentation and counting of nuclei in breast cancer pathology images. In the development of the algorithm, we proposed two novel methods: an adaptive-sized hybrid neural network for the automatic segmentation of nuclei, insulin-like growth factor-II messenger RNAs and other structures, and the combined use of both the focused gradient filter and the watersheds algorithm for segmentation of overlapped nuclei.

Several methods have been used to obtain complete MR images from a reduced set of measured encodes. All of these techniques use some sort of a priori information to complete the incomplete data set. These techniques have had varying degrees of success in a variety of applications: imaging contrast uptake, interventional procedures, spectroscopy, fMRI contrast changes, and routine scans of a given population. In this paper we make three points: First, significant effort has been made toward finding the set of phase encodes that minimizes the expected L₂ norm of the encoding error. We show that in experiments many sets of phase encodes come close to attaining the Karhunen-Loeve (K-L) limit. However, second, we show that other image quality metrics result in much more pleasing images that have much better detail than images where the L₂ norm is optimized. For this purpose we use best basis algorithms to find a local trigonometric and wavelet packet bases which optimizes several cost functions including the L_p norms and entropy; the L_0.25 norm gave the most pleasing results. Lastly, we show that although ringing produced by undersampling cannot be eliminated it can be effectively controlled when the general shape of the object is known. Ringing renders techniques like SLIM impractical. We are able to control ringing by fitting the reconstructed signal to a piecewise monotonic function. The larger peaks are kept and the fast oscillations characteristic of ringing are eliminated. The signal can be pulled out of noise or fast oscillations when the signal energy is only one quarter of the noise energy.

We have developed a region-based image processing method to enhance selective radiodense regions on digital radiographs. We employ a wavelet filtering technique to locate the radiodense regions-of-interest and then apply different degrees of enhancement procedure to them. The enhancement procedure is based on an unsharp masking technique controlled by a set of sigmoidal functions. The method was tested on computed chest radiographs to improve the visualization of the mediastinum and radiodense spine areas. The enhanced chest images showed improved visualization in the mediastinum area, and the visibility of vascular structures which were obscured by the diaphragm and mediastinum was improved. To demonstrate the method's potential in other medical image processing tasks, we applied it to cervical spine images. The processed cervical spine images also showed better visualization of the seventh cervical vertebrae and the first thoracic vertebrae in the high radiodense area caused by the superimposition of the patient's shoulder tissue over these regions of interest.

The sensitivity of mammography for the detection of breast cancer is decreased in the radiodense breast. Storage phosphor digital radiographic systems have a wider latitude of exposure than conventional mammographic screen film systems. By using low resolution histogram equalization one can produce a mammographic image of the breast that retains the high frequency information that defines the edges of microcalcifications, architectural distortion and some masses but which, at the same time, allows one to look through into regions of increased breast radiodensity and identify microcalcifications within them. This paper demonstrates the effect of this special form of image processing.

We are developing a computer program for automated classification of clustered microcalcifications associated with benign and malignant processes in mammograms. Accurate classification of microcalcifications into benign and malignant groups would help improve diagnosis sensitivity as well as reduce the number of unnecessary biopsies. In this study we investigate the effectiveness of several image features and the proposed neural network, the Trend-Oriented Radial Basis Function (TRBF). Our database is composed of 47 benign and 81 malignant region of interest (ROI) images, a total of 128 ROIs, which are selected from 50 micron by 50 micron digitized whole mammograms manually. Each 256 by 256 pixel ROI image contains clustered microcalcifications. First we extract 16 image features which are calculated from a binarized microcalcifications image and its two morphological dilation images. These features are based on 3 morphological criteria: (1) size and/or shape of the calcifications, (2) size or shape of the 'cluster,' and (3) number of microcalcifications. Secondly, we apply Karhunen-Loeve (K-L) expansion to 16 dimension feature space in order to reduce the dimension of the problem. Next we select two dimensional K-L features, which are outcomes of K-L expansion, through a calculation of the Euclidean distance measure. Finally we classify them based on two dimensional K-L features using the proposed neural network. The proposed TRBF neural network has three layers and a new learning algorithm. Its cost function for the learning process consists of a sum of squares error and a sum of inverse RBFs width. Through the learning process, the center and the width of each RBF and the weights between hidden and output layers are decided. The performances of the K-L features and the TRBF neural network are evaluated through the round-robin method in which one sample is tested after learning based on the rest of 127 samples. The two dimensional K-L features are more distinguishable than the raw two dimensional feature combination. The proposed TRBF network is able to define the trend of the distribution better than former RBF networks. According to the receiver operating characteristic (ROC) curve, this system indicates a better performance than one trained radiologist.

The paper presents a method for automatic localization and segmentation of white blood cells (WBCs) with color images to develop an efficient automated leukocyte counter by using pattern recognition-based slide readers. The segmentation techniques consist of the following steps. On the first a smear image acquired at the low magnification. The next is extraction of WBC nuclei by chromatic properties and image mapping. After this the cells clustered according to the distances between them and regions of interest (ROI) determined. Image of ROI captured at the high magnification and its validity checked. Then nucleus segments extracted and grouped into prospective cells. The detection of blood cells is based on the intensity of G image plane and the balance between G and B intensity of the nuclei. A cytoplasm region approximated by a circle area around the nucleus center. Finally, the cytoplasm area cleaned considering a priori knowledge of background color and possible cell occlusions. The result of the segmentation is presented in the form of a cell location list and image template in which every pixel is assigned to a label such as Background, Cytoplasm, Nucleus, Hole, etc. The proposed technique has yielded correct segmentation of complex image scenes for blood smears prepared by ordinary manual staining methods in 99% of tested images.

In this article we present a new solution for the problem of contrast enhancement. The method is based on the decomposition of the original image into its primitive components. In a first stage a multiresolution representation of the image is computed yielding transform coefficients proportional to the maximum of the local gradient at a specific scale and at a specific position in the image. As an image detail exists at several scales, the transform coefficients corresponding to the same image primitive are grouped. Contrast is enhanced by groupwise non-linear amplification of the transform coefficients. An inverse transform is then applied to the modified coefficients, resulting in a uniformly contrast-enhanced image without artifacts. Preliminary results demonstrate the effectiveness of this method. It could be used in a wide range of applications.

Surface (e.g., skin or cortex) based methods to register SPECT and MR images are well known and widely used. Such methods have the disadvantage of needing some kind of segmentation to obtain the surface, which is often a high-level task and possibly error-prone. Also, when reducing the gray-valued images to surfaces, i.e., binary structures, valuable information may be lost in the process. We propose to use the same surfaces, but in a fuzzy way, i.e., we determine the 'surfaceness,' thus retaining more information and avoiding binary segmentation. Such surfaceness can be computed using various techniques. We found a simple morphological approach (which is easy to implement and has very low computational complexity) to be quite sufficient. The resultant surfaceness images are subsequently registered by optimizing the cross-correlation value. The optimization is done using a hierarchical approach to cut the number of required computer operations down to an acceptable level.

A multi-resolution unsharp masking (USM) technique is developed for image feature enhancement in digital mammogram images. This technique includes four processing phases: (1) determination of parameters of multi-resolution analysis (MRA) based on the properties of images; (2) multi-resolution decomposition of original images into sub-band images via wavelet transformation with perfect reconstruction filters; (3) modification of sub-band images with adaptive unsharp masking technique; and (4) reconstruction of image from modified sub- band images via inverse wavelet transformation. An adaptive unsharp masking technique is applied to the sub-band images in order to modify the pixel values based on the edge components at various frequency scales. Smoothing and gain factor parameters, employed in the unsharp masking, are determined according to the resolution, frequency, and energy content of the sub-band images. Experimental results show that this technique is able to enhance the contrast of region of interest (microcalcification clusters) in mammogram image.

Box counting (BC) is widely used to estimate the fractal dimension (fd) of medical images on the basis of a finite set of pixel data. The fd is then used as a feature to discriminate between healthy and unhealthy conditions. We show that BC is ineffective when used on small data sets and give examples of published studies in which researchers have obtained contradictory and flawed results by using BC to estimate the fd of data-limited medical images. We present a new method for estimating fd of data-limited medical images. In the new method, fractal interpolation functions (FIFs) are used to generate self-affine models of the underlying image; each model, upon discretization, approximates the original data points. The fd of each FIF is analytically evaluated. The mean of the fds of the FIFs is the estimate of the fd of the original data. The standard deviation of the fds of the FIFs is a confidence measure of the estimate. The goodness-of-fit of the discretized models to the original data is a measure of self-affinity of the original data. In a test case, the new method generated a stable estimate of fd of a rib edge in a standard chest x-ray; box counting failed to generate a meaningful estimate of the same image.

Assessment of cardiac function using imaging techniques requires accurate identification of borders. We have developed a semi-automatic method for boundary detection of echocardiographic images. Our method combines morphological image analysis techniques with a second derivative operator, namely Laplican of Gaussian. Morphological filtering is used to reduce noise and increase contrast of the image. Results of second order filtering are used to locate edge points, as well as incorporate with morphological segmentation to extract the contour. We validate the results by comparing the computer generated contours to the contours manually outlined by an expert on the data sets. The average of the root mean square errors of the distances is 2.56 pixels (sigma equals 1.21 pixels) for the endocardial borders. The correlation coefficient between the areas enclosed is 0.99. Thus, by using morphological image analysis in echocardiographic images, the overall procedure is simple, efficient and produces encouraging results.

Segmentation is an important step in image analysis. Thresholding is one of the most important approaches. There are several difficulties in segmentation, such as automatic selecting threshold, dealing with intensity distortion and noise removal. We have developed an adaptive segmentation scheme by applying the Central Limit Theorem in regression. A Gaussian regression is used to separate the distribution of background from foreground in a single peak histogram. The separation will help to automatically determine the threshold. A small 3 by 3 widow is applied and the modal of the local histogram is used to overcome noise. Thresholding is based on local weighting, where regression is used again for parameter estimation. A connectivity test is applied to the final results to remove impulse noise. We have applied the algorithm to x-ray angiogram images to extract brain arteries. The algorithm works well for single peak distribution where there is no valley in the histogram. The regression provides a method to apply knowledge in clustering. Extending regression for multiple-level segmentation needs further investigation.

Three-dimensional (3D) image data generated by modalities such as CT and MRI can provide a physician with insight. Analysis of such data has traditionally been done using time- consuming, error-prone manual techniques. Automating this analysis can improve the consistency of the results and reduce the amount of user interactive time required. Automated analysis, however, carries with it a tremendous computational burden. To address this problem, we have designed a distributed processing environment for the rapid processing of 3D radiological images. Our system enables sequences of image processing operations (filtering, morphology, segmentation, etc.) to be linked together. These sequences are then rapidly carried out by distributing the computational load across a network of workstations and a massively parallel computer, using the PVM protocol. A graphical user interface facilitates system operation, making the low-level parallel processing issues transparent to the user.

The filtered-backprojection (FBP) algorithm and statistical model based iterative algorithms such as the maximum likelihood (ML) reconstruction are the two major classes of tomographic reconstruction method. The FBP method is widely used in clinical setting while iterative methods have attracted research interests in the past decade. In this paper we study the performance of the FBP and the ML methods using simulated projection data. The results indicate that the best image that the FBP or the ML algorithm can generate is the compromise of image smoothness and sharpness. The filter cutoff frequency for the FBP algorithm or the number of iterations for the ML algorithm has to be selected carefully.

Shape is a powerful descriptor in identifying and describing objects in images, but is largely unused as a tool for object segmentation and computer-aided image analysis. We extract shape descriptors of objects, 'cores,' in computer images derived from a variety of sources, and prove their utility in important biomedical applications, including segmentation of the hippocampus and cerebral ventricles from MR images, and discrimination between normal and abnormal nerve axons in photomicrographs.

Computer-aided diagnosis techniques have been proposed as second opinion providers in digital mammography. This paper considers a new method of presenting CAD output to the radiologist. Instead of superimposing detected pixels or arrows on the mammogram, we adaptively enhance the most suspicious regions according to the weight indicated by the test statistic at the detector output. In so doing, CAD false positives promise to be less obtrusive to the viewer, and lesions missed by CAD (false negatives) may still be detected by the radiologist. In our method the entire mammogram is enhanced to some (spatially varying) degree. Enhancement is realized by applying nonlinear operators to wavelet coefficients computed at multiple scales. We combine this technique with the results of our previous wavelets-based CAD algorithm for detecting microcalcifications.

This paper reports recent advances in the development of an automated approach to neuropsychological testing. High performance image analysis algorithms have been developed as part of a convenient and non-invasive computer-based system to provide an objective assessment of patient responses to figure-copying tests. Tests of this type are important in determining the neurological function of patients following stroke through evaluation of their visuo-spatial performance. Many conventional neuropsychological tests suffer from the serious drawback that subjective judgement on the part of the tester is required in the measurement of the patient's response which leads to a qualitative neuropsychological assessment that can be both inconsistent and inaccurate. Results for this automated approach are presented for three clinical populations: patients suffering right hemisphere stroke are compared with adults with no known neurological disorder and a population comprising normal school children of 11 years is presented to demonstrate the sensitivity of the technique. As well as providing a more reliable and consistent diagnosis this technique is sufficiently sensitive to monitor a patient's progress over a period of time and will provide the neuropsychologist with a practical means of evaluating the effectiveness of therapy or medication administered as part of a rehabilitation program.

The aim of this work is to increase the precision of the computations involved in brain tumor stereotactic radiosurgery. It is proposed to apply the newest algorithm of contour segmentation to the contours of the MR/CT slices. It is shown that this algorithm improves significantly the accuracy and speed of the classical contour segmentation algorithms and produces a one-to-one transformation between the Euclidean space and the discrete pixel space. For the distance computations involved in dose distribution computation a subpixel precision is obtained. The segmentation algorithm computes the exact equation of the discrete lines forming a pixel contour of an object (the skull, a tumor, etc.). Since it actually computes the equation of the lines forming the contour and not an approximated line like most algorithms, the algorithm is reversible. From the segmented contour the pixel contour can be exactly reconstructed without loss or displacement of any pixel. The segmentation algorithm is quick since it works in linear time. The present application involves the computation of dose distribution for stereotactic radiosurgery. The intersection between the ray and the skull, the sinus cavities and the tumor can be computed with subpixel accuracy. This obviously improves the dose distribution computation. There are many other applications for this algorithm, for example segmentation, slice reconstruction or simply better rendering of anatomical information, etc. The advantage of this segmentation algorithm over classical approaches lies not only in the results presented here, but also in the fact that a 3D extension should be available soon. This strongly suggests the building of a real 3D planning system, simplifies the inverse planning problem and increases the precision of the computations involved in stereotactic radiosurgery.

The problem of mis-registration of 3D datasets is addressed for MR imaging of the breast. Three methods are examined and qualitative evidence of their performance is presented. A method of constrained principal axis determination is compared with a correlation method, and an orthogonal polynomial approach is extended into three dimensions. The most effective methods still remain relatively time consuming, and for the correlation method, parallel distributed computing has been applied to reduce the time-to-solution.

In this investigation, automatic image segmentation is carried out on magnetic resonance image (MRI). A novel technique based on the maximum minimum measure is devised. The measure is improved by combining the smoothing and counting processes, and then normalizing the number of maximum and minimum positions over the region of interest (ROI). Two parameters (MM_H and MM_V) are generated and used for the segmentation. The technique is tested on some brain MRIs of a human male from the Visible Human Project of the National Library of Medicine, National Institutes of Health, USA. Preliminary results indicate that the maximum minimum measure can provide effective parameters for human tissue characterization and image segmentation with an added advantage of faster computation.

Breast cancer is a leading cause of death among women. Mammography is the best established procedure for breast cancer screening and diagnosis. However, mammograms are hard to interpret even by highly-qualified radiologists, especially in cancers at their early stages. In this paper, we analyze the effectiveness of our adaptive neighborhood contrast enhancement (ANCE) technique in increasing breast cancer diagnosis sensitivity. Two hundred and thirty- three film-screen mammograms of twenty-eight interval cancer patients and six benign control cases were digitized with high-resolution of about 4096 by 2048 by 10-bit pixels and then processed with our ANCE method. Unprocessed and processed digitized mammograms as well as the original films were presented to three reference radiologists for a receiver operating characteristic (ROC) evaluation. The results show that the radiologists' performance with the ANCE-processed images is the best among the three sets of images (original, digitized, and enhanced) in terms of area under the ROC curve, and that diagnostic sensitivity is improved by our ANCE algorithm. All of the 19 interval cancer cases not detected with original films of earlier mammography tests were diagnosed as malignant with the corresponding ANCE- processed versions, while only one of six benign patients initially labeled correctly with the original mammograms was interpreted as malignant after enhancement. This study demonstrates the potential of diagnostic performance improvement for early detection of breast cancer with proper digital image enhancement.

This paper is devoted to a statistical analysis of textures from two codings. This analysis discriminates textures: a classification is built, not directly upon texture images, but upon a compressed information, which is created from each texture and composed of two coding images. The principle of this analysis is as follows: a texture images set is used. Each texture, composed of 256 gray levels, is encoded by two coding images of 15 colors. Some Kolmogorov-Smirnov's tests are carried out on combinations of two coding images in view to get a first discrimination. In the same time, from each coding image, co-occurrence parameters are computed. These parameters are used, with the previous discrimination, to get classifications. These classifications are compared to another one made only with co- occurrence parameters directly computed from a basis-textures set. In conclusion, we consider advantages and drawbacks of our approach and perspectives for the future.

A totally automatic scheme for segmenting brain from extracranial tissues and to classify all intracranial voxels as CSF, gray matter (GM), white matter (WM), or abnormality such as multiple sclerosis (MS) lesions is presented in this paper. It is observed that in MR head images, if a tissue's intensity values are normalized, its relationship to the other tissues is essentially constant for a given type of image. Based on this approach, the subcutaneous fat surrounding the head is normalized to classify other tissues. Spatially registered 3 mm MR head image slices of T1 weighted, fast spin echo [dual echo T2 weighted and proton density (PD) weighted images] and fast fluid attenuated inversion recovery (FLAIR) sequences are used for segmentation. Subcutaneous fat surrounding the skull was identified based on intensity thresholding from T1 weighted images. A multiparametric space map was developed for CSF, GM and WM by normalizing each tissue with respect to the mean value of corresponding subcutaneous fat on each pulse sequence. To reduce the low frequency noise without blurring the fine morphological high frequency details an anisotropic diffusion filter was applied to all images before segmentation. An initial slice by slice classification was followed by morphological operations to delete any brides connecting extracranial segments. Finally 3-dimensional region growing of the segmented brain extracts GM, WM and pathology. The algorithm was tested on sequential scans of 10 patients with MS lesions. For well registered sequences, tissues and pathology have been accurately classified. This procedure does not require user input or image training data sets, and shows promise for automatic classification of brain and pathology.

The study of in vitro cell motility has received a great deal of attention for many years. All former studies typically imply the use of a microscope equipped with a camera, a digitizer board and a computer that process the frames to evaluate cell motility. The present paper reports an improved device incorporating the typical components of in vitro cell motility analyzers, which also makes possible the evaluation of clustered cell motility, an important variable for assessing aggressiveness in cancer models. Two methods have been developed. The first avoids the segmentation step and thus provides a measurement of the motility of the whole cluster considered as a single entity. The second consists of a clustered cell tracking software developed in our laboratory. The segmentation process is performed by means of an algorithm based on watershed transformation applied to a modified gradient image. This gradient modification consists of a new algorithm using the correlation between the frames. As a result, the cells are extracted from the background and can be tracked through the sequence. A set of five experiments was performed on cell colonies from the PC-3 human prostate cancer in vitro model colonies in order to illustrate the features of the device.

The detection of abnormalities in screening mammograms requires the highest possible image quality standards, since such abnormalities lie at the limit of physical detection. Image quality is currently assessed in the UK by imaging test objects most of which use a scoring system to gauge the visibility of details. This assessment is clearly subjective and open to inconsistencies over time and between observers. Image processing techniques were used to investigate the feasibility of automating the assessment of image quality. Test object radiographs were digitized using a Lumysis laser digitizer and a dedicated assessment procedure was developed to detect features within the image. Two methods were developed which identified and located test object details and were compared to human observers scoring the same test object radiograph. Contrast detail curves were used to compare all performances. The computer methods compared well with the human observers when identifying small details of less than 1 mm but were less successful in locating larger diameter details.

We describe a database-aided system for diagnosing mammograms with automatic selection of features (i.e., attributes of objects or textures in images). A database of digitized images of calcifications and textures assists the diagnosis and reduces the negative biopsy rate. In our system mammograms are filtered to enhance the calcifications and textures; image features are automatically extracted and stored. Genetic selection and visual mapping of features permits global visualization of the entire database or selected portions of the databases for interaction with mammographers. We demonstrate that this system is able to cluster and retrieve visually and diagnostically similar mammograms.

This report describes a computerized system for assisting with the image analysis of digitized spine x-ray films. It was developed for vertebral height assessment (morphometry), a procedure widely used in research and clinical trials of spinal osteoporosis. Good image presentation is important for effective, interactive analysis of x-ray images, and was achieved by using a 12-bit color map for the image display. This provides virtually instantaneous updating of image display parameters even with large (4 - 10 MB) image files. The software has both semi-automated and automated features, and is integrated with a relational database for the storage and retrieval of images and textual information. It contains an on-line tutorial and facilities for later review and editing of the analyzed data. An active contour model for obtaining additional indices of vertebral fracture has been incorporated, and is undergoing further development. We discuss the organization of the analysis software, the database and on-line tutorial, and present some of the details of the implementation of the active contour model.

A system is being developed to support examination of inflated colons with helical computed tomography (CT). These CT scans are typically reconstructed with sampling intervals smaller than 1 cubic mm. Significant colon polyps (those larger than 5 mm in diameter) are readily recognized. However, direct interpretation of the cross sectional images is prohibitively difficult because a scan of the colon consists of several hundred axial tomograms. Without advanced image manipulation tools, interpretation of the colon's complex three dimension shape from these cross sections alone is nearly impossible. New analysis software is addressing this difficulty.

A new approach, a pyramid template matching as guided by a snake, is developed for the extraction of pelvic features in both portal and simulation images. Initially, the treatment field edge was extracted using a Canny edge detector. A template modeled using a polynomial function for the typical pelvic structures was used as the initial approximation for the anteroposterior (AP) pelvic brim. Several energies were defined to search the actual bony structures as guided by the snake in a larger range. In this study, a double-snake model coupled with a spring was developed to define the external constraint force. The image force was calculated from various processed images using different edge detection algorithms. The criterion used for search termination was to find the locations where the overall energy was at its minimum. The result of the initial search was fit using a polynomial function as the second approximation for the pelvic bony structure. Snake searching technique was repeated in a smaller range around the initial identified features for fine search of pelvic bony structure. This technique has shown to be very promising for extracting pelvic features.

The purpose of our study was to characterize solitary pulmonary nodules (SPN) as benign or malignant based on pattern classification techniques using size, shape, density and texture features extracted from HRCT images. HRCT images of patients with a SPN are acquired, routed through a PACS and displayed on a thoracic radiology workstation. Using the original data, the SPN is semiautomatically contoured using a nodule/background threshold. The contour is used to calculate size and several shape parameters, including compactness and bending energy. Pixels within the interior of the contour are used to calculate several features including: (1) nodule density-related features, such as representative Hounsfield number and moment of inertia, and (2) texture measures based on the spatial gray level dependence matrix and fractal dimension. The true diagnosis of the SPN is established by histology from biopsy or, in the case of some benign nodules, extended follow-up. Multi-dimensional analyses of the features are then performed to determine which features can discriminate between benign and malignant nodules. When a sufficient number of cases are obtained two pattern classifiers, a linear discriminator and a neural network, are trained and tested using a select subset of features. Preliminary data from nine (9) nodule cases have been obtained and several features extracted. While the representative CT number is a reasonably good indicator, it is an inconclusive predictor of SPN diagnosis when considered by itself. Separation between benign and malignant nodules improves when other features, such as the distribution of density as measured by moment of inertia, are included in the analysis. Software has been developed and preliminary results have been obtained which show that individual features may not be sufficient to discriminate between benign and malignant nodules. However, combinations of these features may be able to discriminate between these two classes. With additional cases and more features, we will be able to perform a feature selection procedure and ultimately to train and test pattern classifiers in this discrimination task.

This paper describes a computerized system of tumor detection for lung cancer diagnosis. Through the ten years study, we developed some key algorithms for computer-aided diagnosis. The most important algorithm is a filter to detect suspicious shadows of tumor on a plain chest x-ray image. The filter is named directional contrast filter for nodule (DCF-N). The DCF-N is highly sensitive to ambiguous shadows such as malignant tumors. And we developed a rule- based system to eliminate false-positive shadows. In our study, the system was effective to eliminate shadows of blood vessels and ribs which were primary groups of false-positives. Our current research is focusing on the development of the automatic tumor detection system for lung cancer examination by using CT images. In this paper, we discuss whether the system for plain chest x-ray images can apply to spiral CT images within a malignant tumor, which are reconstructed at 2 mm or 5 mm slice thickness. About a trial case, although the DCF-N can detect the malignant tumor, some false-positives are also detected. As to the analysis of the shadows, which are detected by the DCF-N, the major false-positives are blood vessels shadows. Therefore, the rule to eliminate blood vessels shadows in the rule-base for plain x- ray images is also effective for the CT images.