In this paper we present preliminary results to automatically segment multiple sclerosis (MS) lesions in multispectral magnetic resonance datasets using support vector machines (SVM). A total of eighteen studies (each composed of T1-, T2-weighted and FLAIR images) acquired from a 3T GE Signa scanner was analyzed. A neuroradiologist used a computer-assisted technique to identify all MS lesions in each study. These results were used later in the training and testing stages of the SVM classifier. A preprocessing stage including anisotropic diffusion filtering, non-uniformity intensity correction, and intensity tissue normalization was applied to the images. The SVM kernel used in this study was the radial basis function (RBF). The kernel parameter (γ) and the penalty value for the errors were determined by using a very loose stopping criterion for the SVM decomposition. Overall, a 5-fold cross-validation accuracy rate of 80% was achieved in the automatic classification of MS lesion voxels using the proposed SVM-RBF classifier.

Age-Related Macular Degeneration (ARMD) is the leading cause of irreversible visual loss among the elderly in the US and Europe. A computer-based system has been developed to provide the ability to track the position and margin of the ARMD associated lesion; drusen. Variations in the subject's retinal pigmentation, size and profusion of the lesions, and differences in image illumination and quality present significant challenges to most segmentation algorithms. An algorithm is presented that first classifies the image to optimize the variables of a mathematical morphology algorithm. A binary image is found by applying Otsu's method to the reconstructed image. Lesion size and area distribution statistics are then calculated. For training and validation, the University of Wisconsin provided longitudinal images of 22 subjects from their 10 year Beaver Dam Study. Using the Wisconsin Age-Related Maculopathy Grading System, three graders classified the retinal images according to drusen size and area of involvement. The percentages within the acceptable error between the three graders and the computer are as follows: Grader-A: Area: 84% Size: 81%; Grader-B: Area: 63% Size: 76%; Grader-C: Area: 81% Size: 88%. To validate the segmented position and boundary one grader was asked to digitally outline the drusen boundary. The average accuracy based on sensitivity and specificity was 0.87 for thirty four marked regions.

Principal Component Analysis of sets of temporal shape sequences renders eigenvariations of shape/motion, including typical normal and pathological endocardial contraction patterns. A previously developed Active Appearance Model for time sequences (AAMM) was employed to derive AAMM shape coefficients (ASCs) and we hypothesized these would allow classification of wall motion abnormalities (WMA). A set of stress echocardiograms (single-beat 4-chamber and 2-chamber sequences with expert-verified endocardial contours) of 129 infarct patients was split randomly into training (n=65) and testing (n=64) sets. AAMMs were generated from the training set and for all sequences ASCs were extracted and statistically related to regional/global Visual Wall Motion Scoring (VWMS) and clinical infarct severity and volumetric parameters. Linear regression showed clear correlations between ASCs and VWMS. Infarct severity measures correlated poorly to both ASCs and VWMS. Discriminant analysis showed good prediction from low #ASCs of both segmental (85% correctness) and global WMA (90% correctness). Volumetric parameters correlated poorly to regional VWMS. Conclusions: 1)ASCs show promising accuracy for automated WMA classification. 2)VWMS and endocardial border motion are closely related; with accurate automated border detection, automated WMA classification should be feasible. 3)ASC shape analysis allows contour set evaluation by direct comparison to clinical parameters.

Despite the advent of sophisticated image analysis algorithms, most SPECT (Single Photon Emission Computerized Tomography)cerebral perfusion studies are assessed visually, leading to unavoidable and significant inter and intra-observer variability. Here, we present an automatic method for evaluating SPECT studies based on a computerized atlas of normal regional cerebral bloodflow(rCBF). To generate the atlas, normal(screened volunteers)brain SPECT studies are registered with an affine transformation to one of them arbitrarily selected as reference to remove any size and orientation variations that are assumed irrelevant for our analysis. Then a smooth non-linear registration is performed to reveal the local activity pattern displacement among the normal subjects. By computing and applying the mean displacement to the reference SPECT image, one obtain the atlas that is the normal mean distribution of the rCBF(up to an affine transformation difference). To complete the atlas we add the intensity variance with the displacement mean and variance of the activity pattern. To investigate a patient's condition, we proceed similarly to the atlas construction phase. We first register the patient's SPECT volume to the atlas with an affine transformation. Then the algorithm computes the non-linear 3D displacement of each voxel needed for an almost perfect shape (but not intensity)fit with the atlas. For each brain voxel, if the intensity difference between the atlas and the registered patient is higher than normal differences then this voxel is counted as "abnormal" and similarly if the 3D motion necessary to move the voxel to its registered position is not within the normal displacements. Our hypothesis is that this number of abnormal voxels discriminates between normal and abnormal studies. A Markovian segmentation algorithm that we have presented elsewhere is also used to identify the white and gray matters for regional analysis. We validated this approachusing 23 SPECT perfusion studies (^99mTc ECD)selected visually for clear diffuse anomalies (a much more stringent test than "easy" focal lesions detection) and 21 normal studies. A leave-one-out strategy was used to test our approach to avoid any bias. Based on the number of "abnormal" voxels, two simple supervised classifiers were tested:(1)minimum distance-to-mean and (2)Bayesian. A voxel was considered "abnormal" if its P value with respect to the atlas was lower that 0.01(1%). The results show that for the whole brain, a combination of the number of intensity and displacement "abnormal" voxel is a powerful discriminant with a 91% classification rate. If we focus only on the voxels in the segmented gray matter the rates are slighty higher.

We would like to give a work-in-progress report on our ophthalmologic diagnostic software system which performs biomechanically-based muscle volume deformations using MR images. For reconstructing a three-dimensional representation of an extraocular eye muscle, a sufficient amount of high resolution MR images is used, each representing a slice of the muscle. In addition, threshold values are given, which restrict the amount of data used from the MR images. The Marching Cube algorithm is applied to the polygons, resulting in a 3D representation of the muscle, which can efficiently be rendered. A transformation to a dynamic, deformable model is applied by calculating the center of gravity of each muscle slice, approximating the muscle path and subsequently adding Hermite splines through the centers of gravity of all slices. Then, a radius function is defined for each slice, completing the transformation of the static 3D polygon model. Finally, this paper describes future extensions to our system. One of these extensions is the support for additional calculations and measurements within the reconstructed 3D muscle representation. Globe translation, localization of muscle pulleys by analyzing the 3D reconstruction in two different gaze positions and other diagnostic measurements will be available.

Noninvasive imaging of electrical function in the human atria is attained by the combination of data from electrocardiographic (ECG) mapping and magnetic resonance imaging (MRI). An anatomical computer model of the individual patient is the basis for our computer-aided diagnosis of cardiac arrhythmias. Three patients suffering from Wolff-Parkinson-White syndrome, from paroxymal atrial fibrillation, and from atrial flutter underwent an electrophysiological study. After successful treatment of the cardiac arrhythmia with invasive catheter technique, pacing protocols with stimuli at several anatomical sites (coronary sinus, left and right pulmonary vein, posterior site of the right atrium, right atrial appendage) were performed. Reconstructed activation time (AT) maps were validated with catheter-based electroanatomical data, with invasively determined pacing sites, and with pacing at anatomical markers. The individual complex anatomical model of the atria of each patient in combination with a high-quality mesh optimization enables accurate AT imaging, resulting in a localization error for the estimated pacing sites within 1 cm. Our findings may have implications for imaging of atrial activity in patients with focal arrhythmias.

Breast tissue density is one of the most cited risk factors in breast cancer development. Nevertheless, estimates of the magnitude of breast cancer risk associated with density vary substantially because of the inadequacy of methods used in tissue density assessment (e.g., subjective and/or qualitative assessment) and lack of a reliable gold standard. We have developed automated algorithms for quantitatively measuring breast composition from digitized mammograms. The results were compared to objective truth as determined by quantitative measures from breast MR images, as well as to subjective truth as determined by radiologists' readings from digitized mammograms using BI-RAD standards. Higher linear correlation between estimates calculated from mammograms using the methods developed herein and estimates derived from breast MR images demonstrates that the mammography-based methods will likely improve our ability to accurately determine the breast cancer risk associated with breast density. By using volumetric measures from breast MR images as a gold standard, we are able to estimate the adequacy and accuracy of our algorithms. The results can be used for providing a calibrated method for estimating breast composition from mammograms.

Mammographic parenchymal patterns have been shown to be associated with breast cancer risk. Fractal-based texture analyses, including box-counting methods and Minkowski dimension, were performed within parenchymal regions of normal mammograms of BRCA1/BRCA2 gene mutation carriers and within those of women at low risk for developing breast cancer. Receiver Operating Characteristic (ROC) analysis was used to assess the performance of the computerized radiographic markers in the task of distinguishing between high and low-risk subjects. A multifractal phenomenon was observed with the fractal analyses. The high frequency component of fractal dimension from the conventional box-counting technique yielded an Az value of 0.84 in differentiating between two groups, while using the LDA to estimate the fractal dimension yielded an Az value of 0.91 for the high frequency component. An Az value of 0.82 was obtained with fractal dimensions extracted using the Minkowski algorithm.

We conducted an observer performance study using receiver operating characteristic (ROC) methodology to evaluate the effects of computer-aided diagnosis (CAD) on radiologists’ performance for characterization of masses on serial mammograms. The automated CAD system, previously developed in our laboratory, can classify masses as malignant or benign based on interval change information on serial mammograms. In this study, 126 temporal image pairs (73 malignant and 53 benign) from 52 patients containing masses on serial mammograms were used. The corresponding masses on each temporal pair were identified by an experienced radiologist and automatically segmented by the CAD program. Morphological, texture, and spiculation features of the mass on the current and the prior mammograms were extracted. The individual features and the difference between the corresponding current and prior features formed a multidimensional feature space. A subset of the most effective features that contained the current, prior, and interval change information was selected by a stepwise procedure and used as input predictor variables to a linear discriminant classifier in a leave-one-case-out training and testing resampling scheme. The linear discriminant classifier estimated the relative likelihood of malignancy of each mass. The classifier achieved a test Az value of 0.87. For the ROC study, 4 MQSA radiologists and 1 breast imaging fellow assessed the masses on the temporal pairs and provided estimates of the likelihood of malignancy without and with CAD. The average A_z value for the likelihood of malignancy estimated by the radiologists was 0.79 without CAD and improved to 0.87 with CAD. The improvement was statistically significant (p=0.0003). This preliminary result indicated that CAD using interval change analysis can significantly improve radiologists’ accuracy in classification of masses and thereby may increase the positive predictive value of mammography.

We compared computerized methods that incorporate automated lesion characterization and methods for the assessment of the breast parenchymal pattern on mammograms in order to better predict the pathological status of a breast lesion. Computer-extracted mass feature automatically characterized the shape, spiculation, contrast, and margin of each lesion. On the digitized mammogram of the contralateral breast, computer-extracted texture features were automatically extracted to characterize the radiographic breast parenchymal patterns. Three approaches were investigated. A computerized risk-modulated analysis system for mammographic images is expected to improve characterization of lesions by incorporating cancer-risk information into the decision-making process.

We are developing a computerized method that detects suspicious areas on ultrasound images, and then distinguishes between malignant and benign-type lesions. The computerized scheme identifies potential lesions based on expected lesion shape and margin characteristics. All potential lesions are subsequently classified by a Bayesian neural net based on computer-extracted lesion features. The scheme was trained on a database of 400 cases (757 images) - consisting of complex cysts, benign and malignant lesions - and tested on a comparable database of 458 cases (1740 images) including 578 normal images. We investigated the performances of lesion detection and subsequent classification by a Bayesian neural net for two tasks. The first task was the distinction between actual lesions and false-positive (FP) detections, and the second task the distinction between actual malignant lesions and all detected lesion candidates. In training, the detection and classification method obtained an Az value of 0.94 in the distinction of false-positive detections from actual lesions, and an Az of 0.91 was obtained on the testing database. The task of distinguishing malignant lesions from all other detections (false-positives plus all benign type lesions) showed to be more challenging and Az values of 0.87 and 0.81 were obtained during training and testing, respectively. For the testing database, the combined detection and classification scheme correctly identified lesions in 82% (0.45 FP per image) of all the patients, and in 100% (0.43 FP malignancies per image) of the cancer patients.

In this paper, we present preliminary results from a highly sensitive and specific CAD system for mammographic masses. For false positive reduction, the system incorporated features derived from shape, fractal, and channelized Hotelling observer (CHO) measurements. The database for this study consisted of 80 craniocaudal mammograms randomly extracted from USF's digital database for screening mammography. The database contained 49 mass findings (24 malignant, 25 benign). To detect initial mass candidates, a difference of Gaussians (DOG) filter was applied through normalized cross correlation. Suspicious regions were localized in the filtered images via multi-level thresholding. Features extracted from the regions included shape, fractal dimension, and the output from a Laguerre-Gauss (LG) CHO. Influential features were identified via feature selection techniques. The regions were classified with a linear classifier using leave-one-out training/testing. The DOG filter achieved a sensitivity of 88% (23/24 malignant, 20/25 benign). Using the selected features, the false positives per image dropped from ~20 to ~5 with no loss in sensitivity. This preliminary investigation of combining multi-level thresholded DOG-filtered images with shape, fractal, and LG-CHO features shows great promise as a mass detector. Future work will include the addition of more texture and mass-boundary descriptive features as well as further exploration of the LG-CHO.

For radiologists lesion margin appearance is of high importance when classifying breast masses as malignant or benign lesions. In this study, we developed different measures to characterize the margin of a lesion. Towards this goal, we developed a series of algorithms to quantify the degree of sharpness and lobulation of a mass margin. Besides, to estimate spiculation of a margin, features previously developed for mass detection were used. Images selected from the publicly available data set "Digital Database for Screening Mammography" were used for development and evaluation of these algorithms. The data set consisted of 777 images corresponding to 382 patients. To extract lesions from the mammograms a segmentation algorithm based on dynamic programming was used. Features were extracted for each lesion. A k-nearest neighbor algorithm was used in combination with a leave-one-out procedure to select the best features for classification purposes. Classification accuracy was evaluated using the area A_z under the receiver operating characteristic curve. The average test A_z value for the task of classifying masses on a single mammographic view was 0.79. In a case-based evaluation we obtained an A_z value of 0.84.

It has been shown that the presence of a blood oxygen level dependent (BOLD) signal in high-field (3T and higher) fMRI datasets can cause stimulus-correlated registration errors, especially when using a least-squares registration method. These errors can result in systematic inaccuracies in activation detection. The authors have recently proposed a new method to solve both the registration and activation detection least-squares problems simultaneously. This paper gives an outline of the new method, and demonstrates its robustness on simulated fMRI datasets containing various combinations of motion and activation. In addition to a discussion of the merits of the method and details on how it can be efficiently implemented, it is shown that, compared to the standard approach, the new method consistently reduces false-positive activations by two thirds and reduces false-negative activations by one third.

Intra-Operative MR imaging is an emerging tool for image guided (neuro)surgery. Due to the small size of the magnets and the short acquisition time, the images produced by such devices are often subject to distortions. In this work, we show the particular case of images provided by an ODIN device (Odin Medical Technologies, Newton, MA 02458, USA). Such images suffer from geometric distortions and an important bias field in the luminance. In order to simultaneously correct these deformations, we propose to register a preoperative ODIN image with a diagnosis MR high resolution image while compensating the bias field.

We present a new method to perform registration of DT-MRI (Diffusion Tensor Magnetic Resonance Imaging) data. The goal of image registration is to determine the spatial alignment between multiple images of the same or different subjects, acquired intra or inter-modality. Registration of DT-MRI is more complex than for scalar data because it contains additional directional information. The exploitation of DT-MR data for registration should improve the accuracy of image matching for scalar data because the information in DT-MRI is complementary to that contained in standard MR images and thus provides additional cues for matching, which can be used both to test registration quality and improve it. Moreover, developing techniques for spatial normalisation of DT-MR images allows for cross-population studies to be performed using the whole tensor. The novelty of the proposed approach is that it uses the tensor orientation to calculate the registration transformation. We have quantitatively shown that this new algorithm reconstructs some synthetic transformations more closely than current techniques. However, further analysis of our results is necessary to quantify the advantage of our methods more clearly.

This paper presents two non-rigid image registration algorithms: Thirion's Demons method and its spline-based extension, and compares their performance on the task of inter-subject registration of MRI brain images. The methods are designed to be fast and derive their speed from the uncoupling of the correspondence calculation and deformation interpolation procedures, each of which are then amenable to efficient implementation. The evaluation results indicate that this uncoupling does not significantly limit the registration accuracy that can be achieved.

It is clinically important to develop novel approaches to accurately assess early response to chemoprevention. We propose to quantitatively measure changes of breast density and breast vascularity in glandular tissue to assess early response to chemoprevention. In order to accurately extract glandular tissue using pre- and post-contrast magnetic resonance (MR) images, non-rigid registration is the key to align MR images by recovering the local deformations. In this paper, a new registration method has been developed using finite-element deformable sheet-curve models to accurately register MR breast images for extraction of glandular tissue. Finite-element deformable sheet-curve models are coupling dynamic systems to physically model the boundary deformation and image deformation. Specifically, deformable curves are used to obtain a reliable matching of the boundaries using physically constrained deformations. A deformable sheet with the energy functional of thin-plate-splines is used to model complex local deformations between the MR breast images. Finite-element deformable sheet-curve models have been applied to register both digital phantoms and MR breast image. The experimental results have been compared to point-based methods such as the thin-plate-spline (TPS) approach, which demonstrates that our method is of a great improvement over point-based registration methods in both boundary alignment and local deformation recovery.

We investigated the effect of the number of cases included in an image database on development of a computer-aided diagnosis (CAD) scheme for the detection of lung nodules, in terms of the performance of the CAD scheme. A total number of 1000 chest radiographs with nodules was used in this study. All images were divided randomly into subsets consisting of the same number of cases from different sources. The subsets we used in this study were 10 sets of 100 cases, 5 sets of 200 cases, and 2 sets of 500 cases. The entire database and all of the subsets were tested by use of the same CAD scheme, but with different parameter settings for consistency tests. When the sensitivities of the CAD scheme for each subset were kept at a level of 70.0 %, the numbers of false positives per image were 0.1 for 100 cases, 0.6 for 200 cases, 2.9 for 500 cases, and 6.2 for 1000 cases. Therefore, the performance of the CAD scheme in detecting lung nodules was strongly affected by the number of cases used. We conclude that a large-scale image database is needed for reliable evaluation of the performance of CAD.

We have developed various segmentation and analysis methods for the quantification of lung nodules in thoracic CT. Our methods include the enhancement of lung structures followed by a series of segmentation methods to extract the nodule and to form 3D configuration at an area of interest. The vascular index, aspect ratio, circularity, irregularity, extent, compactness, and convexity were also computed as shape features for quantifying the nodule boundary. The density distribution of the nodule was modeled based on its internal homogeneity and/or heterogeneity. We also used several density related features including entropy, difference entropy as well as other first and second order moments. We have collected 48 cases of lung nodules scanned by thin-slice diagnostic CT. Of these cases, 24 are benign and 24 are malignant. A jackknife experiment was performed using a standard back-propagation neural network as the classifier. The LABROC result showed that the Az of this preliminary study is 0.89.

In this paper, we propose a method for reducing false positives in X-ray CT images using ridge shadow extraction techniques and 3D geometric object models. Suspicious shadows are detected by our variable N-quoit (VNQ) filter, which is a type of mathematical morphology filter. This filter can detect lung cancer shadows with the sensitivity over 95[%], but it also detects many false positives which are mainly related to blood vessel shadows. We have developed two algorithms to distinguish lung nodule shadows from blood vessel shadows. In the first algorithm, the ridge shadows, which come from blood vessels, are emphasized by our Tophat by Partial Reconstruction filter which is also a type of mathematical morphology filter. And then, the region of the ridge shadow is extracted using binary distance transformation. In the second algorithm, we propose a recognition method of nodules using 3D geometric lung nodule and blood vessel models. The anatomical knowledge about the 3D structures of nodules and blood vessels can be reflected in recognition process. By applying our new method to actual CT images (37 patient images), a good result has been acquired.

A novel method for the assessment and display of the distribution of emphysema in low-dose helical CT scans has been developed. The automated system segments the lung volume and estimates the degree of emphysema as a function of slice position within the lung. Eighty low-dose (120 kVp, 40 mA) high-resolution (2.5 mm slice thickness) CT scans were randomly selected from our lung cancer screening program. Three emphysema assessments were performed on each scan: the traditional method of averaging the degree of emphysema on four pre-selected CT slices, the total volumetric percentage of emphysema, and a graphical display of emphysema burden as a function of slice position based on a sliding window algorithm. The traditional four-slice estimates showed a high correlation (0.98) with the total volumetric percentages, yet provided limited spatial information. In those cases with a higher overall percentage of emphysema, the distribution within the lung as quantified by the new method was more skewed than that of less severe cases or normals. Analysis and display of the spatial distribution of emphysema allows for assessment of emphysema burden within each lung zone, which may be useful for quantitating the type of emphysema and the progression of disease over time.

Previously, we presented a Constraint Satisfaction Neural Network (CSNN) to predict the outcome of breast biopsy using mammographic and clinical findings. Based on 500 cases, the study showed that CSNN was able to operate not only as a predictive but also as a knowledge discovery tool. The purpose of this study is to validate the CSNN on a database of additional 1,030 cases. An auto-associative backpropagation scheme was used to determine the CSNN constraints based on the initial 500 patients. Subsequently, the CSNN was applied to 1,030 new patients (358 patients with malignant and 672 with benign lesions) to predict breast lesion malignancy. For every test case, the CSNN reconstructed the diagnosis node given the network constraints and the external inputs to the network. The activation level achieved by the diagnosis node was used as the decision variable for ROC analysis. Overall, the CSNN continued to perform well over this large dataset with ROC area of A_z=0.81±0.02. However, the diagnostic performance of the network was inferior in cases with missing clinical findings (A_z=0.80±0.02) compared to those with complete findings (A_z=0.84±0.03). The study also demonstrated the ability of the CSNN to effectively impute missing findings while performing as a predictive tool.

We investigated a new approach to improve the performance of a computer-aided detection (CAD) scheme in identifying masses depicted on images acquired earlier ("prior"). The scheme was trained using a dataset with simulated mass features. From a database with images acquired during two consecutive examinations, 100 locations matched pairs of malignant mass regions were selected in both the “current” and the most recent “prior” images. While reviewing the current images, mass regions were identified and as a result biopsies were ultimately performed. Prior images were not identified as suspicious by radiologists during the original interpretation. The same number of false-positive regions was also selected in both current and prior images. The selected regions were then randomly divided into training and testing datasets with 50 true-positive and 50 false-positive regions in each. For each selected region, five features; area, contrast, circularity, normalized standard deviation of radial length, and conspicuity; were computed. The ratios of the average difference of five feature values between current and prior mass regions in the training datasets were also computed. Multiplying these ratios by the computed values in current mass regions, we generated a new dataset of simulated features of “prior” mass regions. Three artificial neural networks (ANN) were trained. ANN-1 and ANN-2 were trained using training datasets of current and prior regions, respectively. ANN-3 was trained using simulated “prior” dataset. The performance of three ANNs was then evaluated using the testing dataset of prior images. Areas under ROC curves (A_z) were 0.613 ± 0.026 for ANN-1, 0.678 ± 0.029 for ANN-2, and 0.667 ± 0.029 for ANN-3, respectively. This preliminary study demonstrated that one could estimate an average change of feature values over time and "adjust" CAD performance for better detection of masses at an earlier stage.

This paper presents hardware and software procedures for automated cell tracking and migration modeling. A time-lapse microscopy system equipped with a computer controllable motorized stage was developed. The performance of this stage was improved by incorporating software algorithms for stage motion displacement compensation and auto focus. The microscope is suitable for in-vitro stem cell studies and allows for multiple cell culture image sequence acquisition. This enables comparative studies concerning rate of cell splits, average cell motion velocity, cell motion as a function of cell sample density and many more. Several cell segmentation procedures are described as well as a cell tracking algorithm. Statistical methods for describing cell migration patterns are presented. In particular, the Hidden Markov Model (HMM) was investigated. Results indicate that if the cell motion can be described as a non-stationary stochastic process, then the HMM can adequately model aspects of its dynamic behavior.

Dynamic contrast enhanced magnetic resonance imaging (DCE-MRI) has emerged as an effective tool to access tumor vascular characteristics. DCE-MRI can be used to characterize noninvasively, microvasculature providing information about tumor microvessel structure and function (e.g., tumor blood volume, vascular permeability, tumor perfusion). However, pixels of DCE-MRI represent a composite of more than one distinct functional biomarker (e.g., microvessels with fast or slow perfusion) whose spatial distributions are often heterogeneous. Complementary to various existing methods (e.g., compartment modeling, factor analysis), this paper proposes a blind source separation method which allows for a computed simultaneous imaging of multiple biomarkers from composite DCE-MRI sequences. The algorithm is based on a partially-independent component analysis, whose parameters are estimated using a subset of informative pixels defining the independent portion of the observations. We demonstrate the principle of the approach on simulated image data set, and we then apply the method to the tissue heterogeneity characterization of breast tumors where spatial distribution of tumor blood volume, vascular permeability, and tumor perfusion, as well as their time activity curves (TACs) are simultaneously estimated.

A limitation of using electron microscopy as a diagnostic tool in virology is the expertise required in analysing and interpreting the images. EM images of different viruses can be very similar in shape. An automated recognition method is proposed in this paper. It is based on radial spectra of higher-order spectral parameters robust to translation, scaling and noise. These features are also roation invariant and can be averaged for a population of viral particles without the need to normalize and align them. They extract symmetry information and are sensitive enough to distinguish viruses that appear nearly circular to the human eye. The method was tested using three such viruses with very similar morphologies - the Adeno, the HAV and the Astro. 70 viral particles of each class from three images were used for training. In the first test, random unseen sets of viral particles form the same images were chosen. In the second test, images of viruses from other sources, where the specimen preparation and the microscope are different, were used to determine the reliability of the system. Both tests have shown high classification accuracy improving rapidly to 100% as the test ensemble grew to 20 particles.

The objective of the current study is to develop an automatic tool to identify bacterial types using computer-vision and statistical modeling techniques. Bacteriophage (phage)-typing methods are used to identify and extract representative profiles of bacterial types, such as the Staphylococcus Aureus. Current systems rely on the subjective reading of plaque profiles by human expert. This process is time-consuming and prone to errors, especially as technology is enabling the increase in the number of phages used for typing. The statistical methodology presented in this work, provides for an automated, objective and robust analysis of visual data, along with the ability to cope with increasing data volumes.

Surface-based representation and classification techniques are studied for hippocampal shape analysis. The goal is twofold: (1) develop a new framework of salient feature extraction and accurate classification for 3D shape data; (2) detect hippocampal abnormalities in schizophrenia using this technique. A fine-scale spherical harmonic expansion is employed to describe a closed 3D surface object. The expansion can then easily be transformed to extract only shape information (i.e., excluding translation, rotation, and scaling) and create a shape descriptor comparable across different individuals. This representation captures shape features and is flexible enough to do shape modeling, identify statistical group differences, and generate similar synthetic shapes. Principal component analysis is used to extract a small number of independent features from high dimensional shape descriptors, and Fisher's linear discriminant is applied for pattern classification. This framework is shown to be able to perform well in distinguishing clear group differences as well as small and noisy group differences using simulated shape data. In addition, the application of this technique to real data indicates that group shape differences exist in hippocampi between healthy controls and schizophrenic patients.

A new algorithm that can be used to automatically recognize and classify malignant lymphomas and lukemia is proposed in this paper. The algorithm utilizes the morphological watershed to extract boundaries of cells from their grey-level images. It generates a sequence of Euclidean distances by selecting pixels in clockwise direction on the boundary of the cell and calculating the Euclidean distances of the selected pixels from the centroid of the cell. A feature vector associated with each cell is then obtained by applying the auto-regressive moving-average (ARMA) model to the generated sequence of Euclidean distances. The clustering measure J3=trace{inverse(Sw^-1)Sm} involving the within (Sw) and mixed (Sm) class-scattering matrices is computed for both cell classes to provide an insight into the extent to which different cell classes in the training data are separated. Our test results suggest that the algorithm is highly accurate for the development of an interactive, computer-assisted diagnosis (CAD) tool.

Quantitative assessment of tree structures is very important for evaluation of airway or vascular tree morphology and its associated function. Our skeletonization and branch-point identification method provides a basis for tree quantification or tree matching, tree-branch diameter measurement in any orientation, and labeling individual branch segments. All main components of our method were specifically developed to deal with imaging artifacts typically present in volumetric medical image data. The proposed method has been tested in a computer phantom subjected to changes of its orientation as well as in a repeatedly CT-scanned rigid plastic phantom. In all cases, our method produced reliable and well positioned centerlines and branch-points.

Thinning is for extracting unit-width skeletons from original objects. Such unit-width skeletons are useful in analyzing tree-structured objects, such bronchi or blood tubes. A tree-structured object could be segmented as a graph since the tails of different branches of the object may be too close and taken as cycles. One possible approach for extracting a tree structure from an original tree-oriented object is to extract a unit-width skeleton, then extract a tree structure from the unit-width skeleton. One major drawback of this approach is that the information of the thickness of each branch is vanished in the first step where the thickness of a branch is important in deciding which voxel should be reduced and which should not. This paper proposes an approach to obtain unit width tree structures from original tree-embedded objects directly through the thinning process.

We propose a technique for suppression of both intra-slice and inter-slice types of motion artifacts simultaneously. Starting from the general assumption of rigid body motion, we consider the case when the acquisition of the k-space is in the form of bands of finite number of lines arranged in a rectilinear fashion to cover the k-space area of interest. We also assume that an averaging factor of at least 2 is desired. Instead of acquiring a full k-space of each image and then average the result, we propose a new acquisition strategy based on acquiring the k-space in consecutive bands having 50% overlap going from one end of the phase encoding direction to the other end. In case of no motion, this overlap can be used as the second acquisition (NEX=2). When motion is encountered, both types motion are reduced to the same form under this acquisition strategy. In particular, detection and correction of motion between consecutive bands result in suppression of both motion types. In this work, this is achieved by utilizing the overlap area to estimate the motion, which is then taken into consideration in further reconstruction (or even acquisition if real-time control is available on the MR system). We demonstrate the accuracy and computational efficiency of this motion estimation approach. Once the motion is estimated, we propose a simple strategy to reconstruct artifact-free images from the acquired data that take into account the possible voids in the acquired k-space as a result of rotational motion between blades.

We present a new framework to estimate and visualize heart motion from echocardiograms. For velocity estimation, we have developed a novel multiresolution optical flow algorithm. In order to account for typical heart motions like contraction/expansion and shear, we use a local affine model for the velocity in space and time. The motion parameters are estimated in the least-squares sense inside a sliding spatio-temporal window. The estimated velocity field is used to track a region of interest which is represented by spline curves. In each frame, a set of sample points on the curves is displaced according to the estimated motion field. The contour in the subsequent frame is obtained by a least-squares spline fit to the displaced sample points. This ensures robustness of the contour tracking. From the estimated velocity, we compute a radial velocity field with respect to a reference point. Inside the time-varying region of interest, the radial velocity is color-coded and superimposed on the original image sequence in a semi-transparent fashion. In contrast to conventional Tissue Doppler methods, this approach is independent of the incident angle of the ultrasound beam. The motion analysis and visualization provides an objective and robust method for the detection and quantification of myocardial malfunctioning. Promising results are obtained from synthetic and clinical echocardiographic sequences.

Tangible solutions to image registration are paramount in longitudinal as well as multi-modal medical imaging studies. In this paper, we introduce tensor scale - a recently developed local morphometric parameter - in rigid image registration. A tensor scale-based registration method incorporates local structure size, orientation and anisotropy into the matching criterion, and therefore, allows efficient multi-modal image registration and holds potential to overcome the effects of intensity inhomogeneity in MRI. Two classes of two-dimensional image registration methods are proposed - (1) that computes angular shift between two images by correlating their tensor scale orientation histogram, and (2) that registers two images by maximizing the similarity of tensor scale features. Results of applications of the proposed methods on proton density and T2-weighted MR brain images of (1) the same slice of the same subject, and (2) different slices of the same subject are presented. The basic superiority of tensor scale-based registration over intensity-based registration is that it may allow the use of local Gestalts formed by the intensity patterns over the image instead of simply considering intensities as isolated events at the pixel level. This would be helpful in dealing with the effects of intensity inhomogeneity and noise in MRI.

Image registration is an important procedure for medical diagnosis. Since the large inter-site retrospective validation study led by Fitzpatrick at Vanderbilt University, voxel-based methods and more specifically mutual information (MI) based registration methods have been regarded as the method of choice for rigid-body intra-subject registration problems. In this study we propose a method that is based on the iterative closest point (ICP) algorithm and a pre-computed closest point map obtained with a slight modification of the fast marching method proposed by Sethian. We also propose an interpolation scheme that allows us to find the corresponding points with a sub-voxel accuracy even though the closest point map is defined on a regular grid. The method has been tested both on synthetic and real images and registration results have been assessed quantitatively using the data set provided by the Retrospective Registration Evaluation Project. For these volumes, MR and CT head surfaces were extracted automatically using a level-set technique. Results show that on these data sets this registration method leads to accuracy numbers that are comparable to those obtained with voxel-based methods.

A quality of image match is usually estimated by measuring image similarity. Unfortunately, similarity measures assess only such transformations that change appearance of the deformed image, and in the case of non-rigid registration the results of the similarity measurement depend on the registration direction. This asymmetric relation leads to registration inconsistency and reduces the quality of registration. In this work we propose a symmetric registration approach, which improves the registration by measuring similarity in both registration directions. The solution presented in this paper is based on the interaction of both images involved in the registration process. Images interact with forces, which are according to the Newton's action-reaction law forming a symmetric relationship. These forces may transform both of the images, although in our implementation one of the images remains fixed. The experiments performed to demonstrate the advantages of the symmetric registration approach involve registration of simple objects, recovering synthetic deformation, and interpatient registration of real images of head. The results show improvements of registration consistency and also indicate the improvement of registration correctness.

In this paper, we examine the problem of image distortion influoroscopy, in particular its effect applications in image-guided surgery that make use of tracking and calibration techniques to relate a point in physical space in the operating room (OR)to the corresponding pointin an intraoperative fluoroscopic image. We call such applications image-enhanced fluoroscopy. In order to derive the relationship between physical space and a fluoroscopic image, two sets of parameters must be known. The first set describes the pose of the fluoroscope at the time when the image was taken; these parameters are usually estimated by rigidly attaching a device to the fluoroscope that is able to be tracked by either an optical or magnetic tracking system present in the OR. The second set of parameters describes the projection model of the fluoroscope itself: this set includes focal length and X-ray source position relative to the image intensifier. In the case that we examine here, these values are estimated using a set of BBs of known geometry placed between the X-ray source and the image intensifier. Because the image intensifier is not perfectly planar, and also because of ambient magnetic fields, the image displayed by the fluoroscope does not match that predicted by a linear projection model. Using the known geometry of the BBs appearing in the image, we may attempt to recover the image that would be given by a linear projection. We call this process dewarping. In this paper, we address the question of the optimal dewarp function. Although a fifth-order polynomial is commonly used, we show that a third-order polynomial is superior. We use real fluoroscope images to demonstrate the potential dangers of using too high a polynomial order for the dewarping process.

All over the world 20% of men are expected to develop prostate cancer sometime in his life. In addition to surgery - being the traditional treatment for cancer - the radiation treatment is getting more popular. The most interesting radiation treatment regarding prostate cancer is Brachytherapy radiation procedure. For the safe delivery of that therapy imaging is critically important. In several cases where a CT device is available a combination of the information provided by CT and 3D Ultrasound (U/S) images offers advantages in recognizing the borders of the lesion and delineating the region of treatment. For these applications the CT and U/S scans should be registered and fused in a multi-modal dataset. Purpose of the present development is a registration tool (registration, fusion and validation) for available CT volumes with 3D U/S images of the same anatomical region, i.e. the prostate. The combination of these two imaging modalities interlinks the advantages of the high-resolution CT imaging and low cost real-time U/S imaging and offers a multi-modality imaging environment for further target and anatomy delineation. This tool has been integrated into the visualization software "InViVo" which has been developed over several years in Fraunhofer IGD in Darmstadt.

This paper describes the application and validation of automatic segmentation of three-dimensional images by non-rigid registration to atlas images. The registration-based segmentation technique is applied to confocal microscopy images acquired from the brains of 20 bees. Each microscopy image is registered to an already segmented reference atlas image using an intensity-based non-rigid image registration algorithm. This paper evaluates and compares four different approaches: registration to an individual atlas image (IND), registration to an average shape atlas image (AVG), registration to the most similar image from a database of individual atlas images (SIM), and registration to all images from a database of individual atlas images with subsequent fuzzy segmentation (FUZ). For each strategy, the segmentation performance of the algorithm was quantified using both a global segmentation correctness measure and the similarity index. Manual segmentation of all microscopy images served as a gold standard. The best segmentation result (median correctness 91 percent of all voxels) was achieved using the FUZ paradigm. Robustness was also the best for this strategy (minimum correctness over all individuals 84 percent). The mean similarity index value of segmentations produced by the FUZ paradigm is 0.86 (IND, 0.81; AVG, 0.84; SIM, 0.82). The superiority of the FUZ paradigm is statistically significant (two-sided paired t-test, P<0.001).

Statistical shape models generally use Principal Component Analysis (PCA) to describe the main directions of shape variation in a training set of example shapes. However, PCA assumes a number of restrictions on the data that do not always hold. In this paper we explore the use of an alternative shape decomposition, Independent Component Analysis (ICA), which does not assume a Gaussian distribution of the input data. Several different methods for performing ICA are available. Three most frequently used methods were tested in order to evaluate their effect on the resulting vectors. In statistical shape models, generally not all the eigenvectors that result from the PCA are used. Vectors de-scribing noise are discarded to obtain a compact description of the data set. The selection of these vectors is based on the natural ordering of the vectors according to the variance in that direction which is inherent to PCA. With ICA, how-ever, there is no natural ordering of the vectors. Four methods for sorting the ICA vectors are investigated. The different ICA-methods yielded highly similar yet not identical results. Vectors obtained with ICA showed localized shape variations, whereas eigenvectors obtained with PCA show global shape variations. From the results of the ordering methods can be seen that PCA is better suited for dimensionality reduction. Of the ordering methods that were tested, the best results were obtained with the ordering according to the locality of the shape variations.

Manual quantitative analysis of cardiac left ventricular function using multi-slice CT is labor intensive because of the large datasets. We present an automatic, robust and intrinsically three-dimensional segmentation method for cardiac CT images, based on 3D Active Shape Models (ASMs). ASMs describe shape and shape variations over a population as a mean shape and a number of eigenvariations, which can be extracted by e.g. Principal Component Analysis (PCA). During the iterative ASM matching process, the shape deformation is restricted within statistically plausible constraints (±3σ). Our approach has two novel aspects: the 3D-ASM application to volume data of arbitrary planar orientation, and the application to image data from another modality than which was used to train the model, without the necessity of retraining it. The 3D-ASM was trained on MR data and quantitatively evaluated on 17 multi-slice cardiac CT data sets, with respect to calculated LV volume (blood pool plus myocardium) and endocardial volume. In all cases, model matching was convergent and final results showed a good model performance. Bland-Altman analysis however, showed that bloodpool volume was slightly underestimated and LV volume was slightly overestimated by the model. Nevertheless, these errors remain within clinically acceptable margins. Based on this evaluation, we conclude that our 3D-ASM combines robustness with clinically acceptable accuracy. Without retraining for cardiac CT, we could adapt a model trained on cardiac MR data sets for application in cardiac CT volumes, demonstrating the flexibility and feasibility of our matching approach. Causes for the systematic errors are edge detection, model constraints, or image data reconstruction. For all these categories, solutions are discussed.

Automatic Left Ventricle (LV) border detection in X-ray angiograms for the quantitative assessment of cardiac function has proven to be a highly challenging task. The main difficulty is segmenting the End Systolic (ES) phase, in which much of the contrast dye has been squeezed out of the LV due to contraction, resulting in poor LV definition. 2D Active Appearance Models (AAMs) have shown utility for segmenting End Diastolic (ED) angiograms, but do not perform satisfactory in individual ES angiograms. In this work, we present a new Multi-view AAM in which we exploit the existing correlation in shape and texture between ED and ES phase to steer the segmentation of both frames simultaneously. Model position and orientation remain independent, whereas appearance statistics are coupled. In addition, an AAM is presented in which the gray-value information of the inner part of the LV is not taken into account. This so-called boundary AAM is applied mainly to enhance local boundary localization performance. Both models are applied in a combined manner and are validated quantitatively. In 61 out of 70 experiments good convergence for both ED and ES segmentation was achieved, with average border positioning errors of 1.86 mm (ED) and 1.93 mm (ES).

The accuracy of discrimination between malignant and benign solid breast masses on ultrasound images may be improved by using computer-aided diagnosis and 3-D information. The purpose of this study was to develop automated 3-D segmentation and classification methods for 3-D ultrasound images, and to compare the classification accuracy based on 2-D and 3-D segmentation techniques. The 3-D volumes were recorded by translating the transducer across the lesion in the z-direction while conventional 2-D images were acquired in the x-y plane. 2-D and 3-D segmentation methods based on active contour models were developed to delineate the mass boundaries. Features were automatically extracted based on the segmented mass shapes, and were merged into a malignancy score using a linear classifier. 3-D volumes containing biopsy-proven solid breast masses were collected from 102 patients (44 benign and 58 malignant). A leave-one-out method was used for feature selection and classifier design. The area Az under the test receiver operating characteristic curves for the classifiers using the 3-D and 2-D active contour boundaries were 0.88 and 0.84, respectively. More than 45% of the benign masses could be correctly identified using the 3-D features without missing a malignancy. Our results indicate that an accurate computer classifier can be designed for differentiation of malignant and benign solid breast masses on 3-D sonograms.

We present the first use of ultrasound to instantiate and register a statistical shape model of bony structures. Our aim is to provide accurate image-guided total hip replacement without the need for a preoperative computed tomography (CT) scan. We propose novel methods to determine the location of the bone surface intraoperatively using percutaneous ultrasound and, with the aid of a statistical shape model, reconstruct a complete three-dimensional (3D) model of relevant anatomy. The centre of the femoral head is used as a further constraint to improve accuracy in regions not accessible to ultrasound. CT scans of the femur from a database were aligned to one target CT scan using a non-rigid registration algorithm. The femur surface from the target scan was then propagated to each of the subjects and used to produce a statistical shape model. A cadaveric femur not used in the shape model construction was scanned using freehand 3D ultrasound. The iterative closest point (ICP) algorithm was used to match points corresponding to the bone Surface derived from ultrasound with the statistical bone surface model. We used the mean shape and the first five modes of variation of the shape model. The resulting root mean square (RMS) point-to-surface distance from ICP was minimised to provide the best fit of the model to the ultrasound data.

A generic approach to building a 3D active appearance model (AAM) for medical image segmentation is presented. Provided a training set of manually segmented images, the model-building procedure is fully automatic. Shape information is obtained from a free-form image registration algorithm. Our AAM is evaluated using the hippocampus as the structure of interest (SOI); the training set consists of 28 manually segmented brain MR scans. The main contributions of this work are: (a) A concept of incorporating the SOI surroundings into the AAM. The concept is also applicable to other medical image based AAMs. (b) A two-step free-form registration procedure (matching the grayscale images first, then matching the segmented images). In this manner, landmark correspondence is improved, and the expert knowledge (i.e., manual segmentations) is less compromised by registration inaccuracies. (c) Two distinct AAM versions are compared: one without and one with statistical texture variation information. Compared to segmentation of an unknown image by registration to a reference image, the main advantage of the AAM is speed: the computation time is brought down from around 5 hours (for free-form deformation computation) to only a few minutes (for optimizing the model parameters), with only a slight degradation in segmentation accuracy.

This paper examines a refinement to probabilistic intensity based tissue segmentation methods, which makes use of knowledge derived from an MRI bias field estimate. Intensity based labeling techniques have employed local smoothness priors to reduce voxel level tissue labeling errors, by making use of the assumption that, within uniform regions of tissue, a voxel should be highly likely to have a similar tissue assignment to its neighbors. Increasing the size of this neighborhood provides more robustness to noise, but reduces the ability to describe small structures. However, when intensity bias due to RF field inhomogeneity is present within the MRI data, local contrast to noise may vary across the image. We therefore propose an approach to refining the labeling by making use of the bias field estimate, to adapt the neighborhood size applied to reduce local labeling errors. We explore the use of a radially symmetric Gaussian weighted neighborhood, and the use of the mean and median of the adapted region probabilities, to refine local probabilistic labeling. The approach is evaluated using the Montreal brainweb MRI simulator as a gold standard providing known gray, white and CSF tissue segmentation. These results show that the method is capable of improving the local tissue labeling in areas most influenced by inhomogeneity. The method appears most promising in its application to regional tissue volume analysis or higher field MRI data where bias field inhomogeneity can be significant.

MR elastography can visualize and measure propagating shear waves in tissue-like materials subjected to harmonic mechanical excitation. This allows the calculation of local values of material parameters such as shear modulus and attenuation. Various inversion algorithms to perform such calculations have been proposed, but they are sensitive to areas of low displacement amplitude (and hence low SNR) that result from interference patterns due to reflection and refraction. A spatio-temporal directional filter applied as a pre-processing step can separate interfering waves so they can be processed separately. Weighted combinations of inversions from such directionally separated data sets can significantly improve reconstructions of shear modulus and attenuation.

Reliably detecting subtle therapy-induced leukoencephalopathy in children treated for cancer is a challenging task due to its nearly identical MR properties and location with unmyelinated white matter. T1, T2, PD, and FLAIR images were collected for 44 children aged 1.7-18.7 (median 5.9) years near the start of therapy for ALL. The ICBM atlas and corresponding apriori maps were spatially normalized to each patient and resliced using SPM99 software. A combined imaging set consisting of MR images and WM, GM and CSF apriori maps were then analyzed with a Kohonen Self-Organizing Map. Vectors from hyperintense regions were compared to normal appearing genu vectors from the same patient. Analysis of the distributions of the differences, calculated on T2 and FLAIR images, revealed two distinct groups. The first large group, assumed normal unmyelinated white matter, consisted of 37 patients with changes in FLAIR ranging from 80 to 147 (mean 117∓17) and T2 ranging from 92 to 217 (mean 144∓28). The second group, assumed leukoencephalopathy, consisted of seven patients with changes in FLAIR ranging from 154 to 196 (mean 171∓19) and T2 ranging from 190 to 287 (mean 216∓33). A threshold was established for both FLAIR (change > 150) and T2 (change > 180).

Trabecular bone (TB) is a network of interconnected struts and plates that constantly remodels to adapt dynamically to the stresses to which it is subjected in such a manner that the trabeculae are oriented along the major stress lines (Wolf's Law). Structural anisotropy can be expressed in terms of the fabric tensor. Next to bone density, TB has been found to be the largest determinant of bone biomechanical behavior. Existing methods, including mean intercept length (MIL), provide only a global statistical average of TB anisotrophy and, generally, require a large sample volume. In this paper, we present a new method, based on the recently conceived notion of tensor scale, which provides regional information of TB orientation and anisotropy. Preliminary evaluation of the method in terms of its sensitivity to resolution and image rotation is reported. The characteristic differences between TB anisotropy computed from transverse and longitudinal sections have been studied and potential applications of the method to in vivo MR imaging are demonstrated. Finally, the ongoing extension of three-dimensional tensor scale in quantitative analysis of tissue morphology is discussed.

In this study we introduce two non-linear structural measures based on the Standard Hough-Transform (SHT) that are applied to high resolution MR-images of human trabecular bone specimens in order to predict biomechanical properties. The results are compared to bone mineral density (BMD) and linear morphometric parameters. Axial MR-images (voxel-size: 117x156x300 mm3) of 33 human femoral and 10 spinal specimens are obtained using a 3D-gradient-echo-sequence. After measurement of BMD by quantitative computed tomography (QCT) all specimens are tested destructively for maximum compressive strength (MCS). The SHT is applied to the binarized and Sobel-filtered images and the peak-value (maxH) and its corresponding bin (posH) of the normalized Hough-spectrum are determined as well as linear measures (apparent bone fraction (app.BV/TV), apparent trabecular separation (app.Tb.Sp), apparent trabecular perimeter per unit area (app.Tb.Perim)). For the spinal [femoral] specimens, R² for MCS vs. maxH is 0.72 (p=0.004) [0.49 (p<0.001)], R² for MCS vs. posH is 0.56 (p=0.013) [0.55 (p<0.001)], and R2 for MCS vs. BMD is 0.43 (p=0.041) [0.72 (p<0.001)]. Correlations of the conventional, linear morphometric parameters and MCS are lower than those for the SHT-based measures or BMD, ranging from 0.20 (p=0.003) for app.BV/TV to 0.46 (p<0.001) for app.Tb.Sp. Prediction of MCS by maxH, posH, or BMD alone is improved by combination with the linear morphometric parameters in a linear regressional model (R² =0.79). In conclusion, the biomechanical strength of human trabecular bone in vitro can effectively be predicted from High-Resolution MR-images by structural measures based on SHT. In the vertebral specimens these are superior to BMD or conventional structural measures in predicting bone strength.

Progressive cerebral atrophy is a physical component of the most common forms of dementia - Alzheimer's disease, vascular dementia, Lewy-Body disease and fronto-temporal dementia. We propose a phenomenological simulation of atrophy in MR images that provides gold-standard data; the origin and rate of progression of atrophy can be controlled and the resultant remodelling of brain structures is known. We simulate diffuse global atrophic change by generating global volumetric change in a physically realistic biomechanical model of the human brain. Thermal loads are applied to either single, or multiple, tissue types within the brain to drive tissue expansion or contraction. Mechanical readjustment is modelled using finite element methods (FEM). In this preliminary work we apply these techniques to the MNI brainweb phantom to produce new images exhibiting global diffuse atrophy. We compare the applied atrophy with that measured from the images using an established quantitative technique. Early results are encouraging and suggest that the model can be extended and used for validation of atrophy measurement techniques and non-rigid image registration, and for understanding the effect of atrophy on brain shape.

A new graph algorithm for the multiscale segmentation of large three-dimensional medical data sets is presented. It is a region-merging segmentation algorithm based on minimizing the Mumford-Shah energy. The Mumford-Shah functional formulation leads to improved segmentation results compared with alternative approaches; and the graph theoretic approach yields improved performance and simplified data structures. Also, the graph algorithm acts on only a subset of the full data set at a given time, allowing its application to large data sets such as whole-body scans. Results on a head MRI data set are presented and compared with a manual segmentation of this data set.

Segmentation of fluid-filled structures, such as the urinary bladder, from three-dimensional ultrasound images is necessary for measuring their volume. This paper describes a system for image enhancement, segmentation and volume measurement of fluid-filled structures on 3D ultrasound images. The system was applied for the measurement of urinary bladder volume. Results show an average error of less than 10% in the estimation of the total bladder volume.

Dynamic single photon emission computed tomography (dSPECT) provides time-varying spatial information about changes of tracer distribution in the body from data acquired using a standard (single slow rotation) protocol. Variations of tracer distribution observed in the images might be due to physiological processes in the body, but may also stem from reconstruction artefacts. These two possibilities are not easily separated because of the highly underdetermined nature of the dynamic reconstruction problem. Since it is expected that temporal changes in tracer distribution may carry important diagnostic information, the analysis of dynamic SPECT images should consider and use this additional information. In this paper we present a segmentation scheme for aggregating voxels with similar time activity curves (TACs). Voxel aggregates are created through region merging based on a similarity criterion on a reduced set of features, which is derived after transformation into eigenspace. Region merging was carried out on dSPECT images from simulated and patient myocardial perfusion studies using various stopping criteria and ranges of accumulated variances in eigenspace. Results indicate that segmentation clearly separates heart and liver tissues from the background. The segmentation quality did not change significantly if more than 99% of the variance was incorporated into the feature vector. The heart behaviour followed an expected exponential decay curve while some variation of time behaviour in liver was observed. Scatter artefacts from photons originating from liver could be identified in long as well as in short studies.

Delineation of structures to irradiate (the tumors) as well as structures to be spared (e.g., optic nerve, brainstem, or eyes) is required for advanced radiotherapy techniques. Due to a lack of time and the number of patients to be treated these cannot always be segmented accurately which may lead to suboptimal plans. A possible solution is to develop methods to identify these structures automatically. This study tests the hypothesis that a fully automatic, atlas-based segmentation method can be used to segment most brain structures needed for radiotherapy plans even tough tumors may deform normal anatomy substantially. This is accomplished by registering an atlas with a subject volume using a combination of rigid and non-rigid registration algorithms. Segmented structures in the atlas volume are then mapped to the corresponding structures in the subject volume using the computed transformations. The method we propose has been tested on two sets of data, i.e., adults and children/young adults. For the first set of data, contours obtained automatically have been compared to contours delineated manually by three physicians. For the other set qualitative results are presented.

Multi slice CT (MSCT) scanners have the advantage of high and isotropic image resolution, which broadens the range of examinations for CT angiography (CTA). A very important method to present the large amount of high-resolution 3D data is the visualization by maximum intensity projections (MIP). A problem with MIP projections in angiography is that bones often hide the vessels of interest, especially the scull and vertebral column. Software tools for a manual selection of bone regions and their suppression in the MIP are available, but processing is time-consuming and tedious. A highly computer-assisted of even fully automated suppression of bones would considerably speed up the examination and probably increase the number of examined cases. In this paper we investigate the suppression (or removal) of bone regions in 3D CT data sets for vascular examinations of the head with a visualization of the carotids and the circle of Willis.

We are investigating magnetic resosance imaging-guided radiofrequency ablation of pathologic tissue. For many tissues, resulting lesions have a characteristic two-boundary appearance featuring an inner region and an outer hyper-intense margin in both contrast enchanced (CE) T₁ and T₂ weighted MR images. We created a twelve-parameter, three-dimensional, globally deformable model with two quadratic surfaces that describe both lesion zones. We present an energy minimization approach to automatically fit the model to a grayscale MR image volume. We applied the automatic model to in vivo lesions (n = 5) in a rabbit thigh model, using CE T₁ and T₂ weighted MR images, and compared the results to multi-operator manually segmented boundaries. For all lesions, the median error was <1.0mm for both the inner and outer regions, values that favorably compare to a voxel width of 0.7 mm. These results suggest that our method provides a precise, automatic approximation of lesion shape. We believe that the method has applications in lesion visualization, volume estimation, image quantification, and volume registration.

The objectives of this paper are to discuss: (1) the development and testing of a new Evolutionary Programming (EP) method to optimally configure Support Vector Machine (SVM) parameters for facilitating the diagnosis of breast cancer; (2) evaluation of EP derived learning machines when the number of BI-RADS and clinical history discriminators are reduced from 16 to 7; (3) establishing system performance for several SVM kernels in addition to the EP/Adaptive Boosting (EP/AB) hybrid using the Digital Database for Screening Mammography, University of South Florida (DDSM USF) and Duke data sets; and (4) obtaining a preliminary evaluation of the measurement of SVM learning machine inter-institutional generalization capability using BI-RADS data. Measuring performance of the SVM designs and EP/AB hybrid against these objectives will provide quantative evidence that the software packages described can generalize to larger patient data sets from different institutions. Most iterative methods currently in use to optimize learning machine parameters are time consuming processes, which sometimes yield sub-optimal values resulting in performance degradation. SVMs are new machine intelligence paradigms, which use the Structural Risk Minimization (SRM) concept to develop learning machines. These learning machines can always be trained to provide global minima, given that the machine parameters are optimally computed. In addition, several system performance studies are described which include EP derived SVM performance as a function of: (a) population and generation size as well as a method for generating initial populations and (b) iteratively derived versus EP derived learning machine parameters. Finally, the authors describe a set of experiments providing preliminary evidence that both the EP/AB hybrid and SVM Computer Aided Diagnostic C++ software packages will work across a large population of patients, based on a data set of approximately 2,500 samples from five different institutions.

Recent developments in imaging technology mean that it is now possible to obtain high-resolution histological image data at multiple wavelengths. This allows pathologists to image specimens over a full spectrum, thereby revealing (often subtle) distinctions between different types of tissue. With this type of data, the spectral content of the specimens, combined with quantitative spatial feature characterization may make it possible not only to identify the presence of an abnormality, but also to classify it accurately. However, such are the quantities and complexities of these data, that without new automated techniques to assist in the data analysis, the information contained in the data will remain inaccessible to those who need it. We investigate the application of a recently developed system for the automated analysis of multi-/hyper-spectral satellite image data to the problem of cancer detection from multispectral histopathology image data. The system provides a means for a human expert to provide training data simply by highlighting regions in an image using a computer mouse. Application of these feature extraction techniques to examples of both training and out-of-training-sample data demonstrate that these, as yet unoptimized, techniques already show promise in the discrimination between benign and malignant cells from a variety of samples.

For the development of computer-aided diagnosis (CAD) systems, a classifier that can effectively differentiate more than two classes is often needed. For example, a detected object on an image may need to be classified as a malignant lesion, a benign lesion, or normal tissue. Currently, a three-class problem is usually treated as a two-stage, two-class problem, in which the detected object is first differentiated as a lesion or normal tissue, and, in the second stage, the lesion is further classified as malignant or benign. In this work, we explored methods for classification of an object into one of the three classes, and compared the three-class approach with the common two-class approach. We conducted Monte Carlo simulation studies to evaluate the dependence of the performance of 3-class classification schemes on design sample size and feature space configurations. A k-dimensional multivariate normal feature space with three classes having different means was assumed. Linear classifiers and artificial neural networks (ANNs) were examined. ROC analysis for the 3-class approach was explored under simplifying conditions. A performance index representing the normalized volume under the ROC surface (NVUS) was defined. Linear classifiers for classification of three classes and two classes were compared. We found that a 3-class approach with a linear classifier can achieve a higher NVUS than that of a 2-class approach. We further compared the performance of an ANN having three or one output nodes with a linear classifier. At large sample sizes, a 3-output-node ANN was basically the same as that of a one-output-node ANN. When the three class distributions had equal covariance matrices and the distances between pairs of class means were equal, the linear classifiers could reach a higher performance for the test samples than the ANN when the design sample size was small; the linear classifier and the ANNs approached the same performance in the limit of large design sample size. However, under complex feature space configurations such as the class means located along a line, the class in the middle was poorly differentiated from the other two classes by the linear classifiers for any dimensionality; the ANN outperformed the linear classifier at all design sample size studied. This simulation study may provide some useful information to guide the design of 3-class classifiers for various CAD applications.

Reliable automated analysis and examination of biomedical images requires reproducible and robust extraction of contained image objects. However, the necessary description of image content as visually relevant objects is context-dependent and determined by parameters such as resolution, orientation, and, of course, the clinical-diagnostic question. Therefore a computer-based approach has to model both examination context and image acquisition as expert knowledge. Generally, static solutions are not satisfying because a change of application will most likely require a redesign of the analysis process. In contrast to non-satisfying statical solution, this paper describes a flexible approach, which allows medical examiners the context-sensitive extraction of sought objects from almost arbitrary medical images, without requiring technical knowledge on image analysis and processing. Since this methodology is applicable to any analysis task on large image sets, it works for general image series analysis as well as image retrieval. The new approach combines classical image analysis with the idea of data mining to close the gap between low abstraction on the technical level and high-level expert knowledge on image content and understanding.

The purpose of the study was to develop and evaluate a content-based image retrieval (CBIR) approach as a computer aid for the detection of masses in screening mammograms. The study was based on the Digital Database for Screening Mammography (DDSM). Initially, a knowledge database of 1,009 mammographic regions was created. They were all 512x512 pixel ROIs with known pathology. Specifically, there were 340 ROIs depicting a biopsy-proven malignant mass, 341 ROIs with a benign mass, and the remaining 328 ROIs were normal. Subsequently, the CBIR algorithm was implemented using mutual information (MI) as the similarity metric for image retrieval. The CBIR algorithm formed the basis of a knowledge-based CAD system. The system operated as follows. Given a databank of mammographic regions with known pathology, a query suspicious mammographic region was evaluated. Based on their information content, all similar cases in the databank were retrieved. The matches were rank-ordered and a decision index was calculated using the query's best matches. Based on a leave-one out sampling scheme, the CBIR-CAD system achieved an ROC area index A_z= 0.87±0.01 and a partial ROC area index _0.90A_z = 0.45±0.03 for the detection of masses in screening mammograms.

In this paper we describe the construction of hierarchical feature clustering and show how to overcome general problems of region growing algorithms such as seed point selection and processing order. Access to medical knowledge inherent in medical image databases requires content-based descriptions to allow non-textual retrieval, e.g., for comparison, statistical inquiries, or education. Due to varying medical context and questions, data structures for image description must provide all visually perceivable regions and their topological relationships, which poses one of the major problems for content extraction. In medical applications main criteria for segmenting images are local features such as texture, shape, intensity extrema, or gray values. For this new approach, these features are computed pixel-based and neighboring pixels are merged if the Euclidean distance of corresponding feature vectors is below a threshold. Thus, the planar adjacency of clusters representing connected image partitions is preserved. A cluster hierarchy is obtained by iterating and recording the adjacency merging. The resulting inclusion and neighborhood relations of the regions form a hierarchical region adjacency graph. This graph represents a multiscale image decomposition and therefore an extensive content description. It is examined with respect to application in daily routine by testing invariance against transformation, run time behavior, and visual quality For retrieval purposes, a graph can be matched with graphs of other images, where the quality of the matching describes the similarity of the images.

The task of segmenting the posterior ribs within the lung fields is of great practical importance. For example, delineation of the ribs may lead to a decreased number of false positives in computerized detection of abnormalities, and hence analysis of radiographs for computer-aided diagnosis purposes will benefit from this. We use an iterative, pixel-based, statistical classification method---iterated contextual pixel classification (ICPC). It is suited for a complex segmentation task in which a global shape description is hard to provide. The method combines local gray level and contextual information to come to an overall image segmentation. Because of it generality, it is also useful for other segmentation tasks. In our case, the variable number of visible ribs in the lung fields complicates the use of a global model. Additional difficulties arise from the poor visibility of the lower and medial ribs. Using cross validation, the method is evaluated on 35 radiographs in which all posterior ribs were traced manually. ICPC obtains an accuracy of 83%, a sensitivity of 79%, and a specificity of 86% for segmenting the costal space. Further evaluation is done using five manual segmentations from a second observer, whose performance is compared with the five corresponding images from the first manual segmentation, yielding 83% accuracy, 84% sensitivity, and 83% specificity. On these five images, ICPC attains 82%, 78%, and 86% respectively.

A novel knowledge-based multi-agent image interpretation system has been developed which is markedly different from previous approaches in especially its elaborate integration of high-level knowledge-based control with low-level image segmentation algorithms. Each agent in this system is responsible for one type of object and cooperates with other agents to come to a consistent overall image interpretation. Cooperation involves communicating hypotheses and resolving conflicts between individual interpretations. Agents have full control over the underlying segmentation algorithms which they dynamically adapt to the image content given knowledge about global constraints, local information and personal beliefs. The system has been applied to IntraVascular Ultrasound (IVUS)images which are segmented by cooperative agents, specialized in lumen, vessel, calcified- plaque, shadow and sidebranch detection. IVUS image sequences from 7 patients were taken and vessel and lumen contours were detected fully automatically. These were compared with expert-corrected semi-automatic contours. Results show good correlation between agents and observer with r=0.84 for the lumen and r=0.92 for the vessel cross-sectional areas(n=1067). The paired difference between agents and observer was 0.13 ± 2.16 mm² for vessel,and -0.14 ± 1.01mm² for lumen cross-sectional areas. These results compare very well with inter-observer variability as reported in the literature.

The problem of vertebrae segmentation in digitized x-ray images is addressed with a hierarchical approach that combines three different methodologies. As a starting point, two customized active shape models are trained on data sets of cervical and lumbar images, respectively. Here, a methodology to include edge information in the gray-level modeling part of the active shape models is developed to increase the representativeness of the model and to improve the chances of finding vertebral boundaries. Active shape models' initialization shortcoming is then addressed by a customized implementation of the Generalized Hough Transform, which provides an estimate of the pose of the vertebrae within target images. Active shape models' shortcoming of lack of local deformation is addressed by a customized implementation of the technique of Deformable Models. In this implementation, an energy minimization approach is employed in which the external energy term is extracted from the training set of images and the internal energy terms control the shape of the template. Segmentation results on data sets of cervical and lumbar images show that the proposed hierarchical approach produces errors of less than 3mm in 75% of the cervical images and 6.4mm in 50% of the lumbar images.

In this paper we present the Interactive Watershed Transform (IWT) for efficient segmentation of multidimensional grayscale images. The IWT builds upon a fast immersion-based watershed transform (WT) followed by a hierarchical organization of the resulting basins in a tree structure. Each local image minimum is represented as an atomic basin at the lowest hierarchy level. The fast WT consists of two steps. First, all image elements are sorted according to their image intensity using a Bucket Sort algorithm. Second, each element is processed exactly once with respect to its neighborhood (e. g., 4, 6, and 8 direct neighbors for 2d, 3d, and 4d transform, respectively) in the specified order. Sort-ing, processing, and tree generation are of order O(n). After computing the WT, one global parameter, the so-called preflooding height, and an arbitrary number of markers are evaluated in real-time to control tree partitioning and basin merging. The IWT has been successfully applied to a large variety of medical images, e. g., for segmentation and volu-metry of neuroanatomic structures as well as bone segmentation, without making assumptions on the objects’ shapes. The IWT combines automation and efficient interactive control in a coherent algorithm while completely avoiding oversegmentation which is the major problem of the classical WT.

A computer simulation of a Charged Fluid was developed to segment objects in medical images. A Charged Fluid conceptually consists of charged particles, each of which exerts a repelling electric force upon the others. In our approach, we treat the image gradients as potential wells where a simulated Charged Fluid flows through. The evolution of a Charged Fluid consists of two different procedures. One allows Charged Fluid elements to advance toward new positions in the simulation along the direction of their total effective forces. The other allows Charged Fluid elements to flow along the contour until an electrostatic equilibrium is achieved. The procedure is repeated until all Charged Fluid elements reside on the boundaries of objects being segmented. Preliminary segmentation results were obtained using this new technique to segment irregular objects in medical images.

Recently one of the authors proposed a reconstruction algorithm, which is theoretically exact and has the truly shift-invariant filtering and backprojection structure. Each voxel is reconstructed using the theoretically minimum section of the spiral, which is located between the endpoints of the PI segment of the voxel. Filtering is one-dimensional, performed along lines with variable tilt on the detector, and consists of five terms. We will present evaluation of the performance of the algorithm. We will also discuss and illustrate empirically the contributions of the five filtering terms to the overall image. A thorough evaluation proved the validity of the algorithm. Excellent image results were achieved even for high pitch values. Overall image quality can be regarded as at least equivalent to the less efficient, exact, Radon-based methods. However, the new algorithm significantly increases efficiency. Thus, the method has the potential to be applied in clinical scanners of the future. The empirical analysis leads to a simple, intuitive understanding of the otherwise obscure terms of the algorithm. Identification and skipping of the practically irrelevant fifth term allows significant speed-up of the algorithm due to uniform distance weighting.

In this paper, we propose a variational framework for tomographic reconstruction of dynamic objects with unknown dynamic models. This is an extension of our previous work on dynamic tomography using curve evolution methods where the shape dynamics are known a priori. We assume the dynamic model of the shape is a parameterized affine transform and propose a variational framework that incorporates information from observed data, intensity dynamics, spatial smoothness prior, and the dynamical shape model. A coordinate descent algorithm based on a curve evolution method is then proposed for the joint estimation of the intensities, object boundary sequences, and the unknown dynamic model parameters. For implementation of the curve evolution and parameter estimation process, we use efficient level set methods.

Calculating digitally reconstructed radiographs (DRRs)is an important step in intensity-based fluoroscopy-to-CT image registration methods. Unfortunately, the standard techniques to generate DRRs involve ray casting and run in time O(n³),where we assume that n is approximately the size (in voxels) of one side of the DRR as well as one side of the CT volume. Because of this, generation of DRRs is typically the rate-limiting step in the execution time of intensity-based fluoroscopy-to-CT registration algorithms. We address this issue by extending light field rendering techniques from the computer graphics community to generate DRRs instead of conventional rendered images. Using light fields allows most of the computation to be performed in a preprocessing step;after this precomputation step, very accurate DRRs can be generated in time O(n²). Using a light field generated from 1,024 DRRs of resolution 256×256, we can create new DRRs that appear visually identical to ones generated by conventional ray casting. Importantly, the DRRs generated using the light field are computed over 300 times faster than DRRs generated using conventional ray casting(50 vs.17,000 ms on a PC with a 2 GHz Intel Pentium 4 processor).

The quasi-exact algorithms developed by Kudo et al. can reconstruct accurate images from data acquired in helical cone-beam configuration. The current formulation of such algorithms, however, prevents their direct application to data acquired in a practical configuration that may be used in helical cone-beam computed tomography (CT) in which the longitudinal axis of the area detector always remains parallel to the longitudinal axis of helical CT system. Interpolation can be used to convert data acquired with the practical helical cone-beam configuration into the form required by the current quasi-exact algorithms. Such an interpolation can reduce spatial resolution in reconstructed images. In this work, we derive a new filtering function that can be used in the quasi-exact algorithms so that they can be used to reconstruct images directly from data acquired with a practical scanning configuration thereby avoiding interpolation. We also performed computer simulation studies, and numerical results from these studies confirm that accurate images can be reconstructed by use of these generalized quasi-exact algorithms. The practical implication of the generalized quasi-exact algorithms is that they can yield images with potentially enhanced spatial resolution by avoiding data interpolation.

We present a novel approach for cardiac motion detection called Cardiac Motion Detection with Complementary Projections (CMDC), which can complement or even supersede the information of the patient's ECG in multi-slice cardiac imaging. High image quality can be obtained using dedicated ECG-correlated cardiac reconstruction algorithms. These algorithms require information about the cardiac motion, usually the simultaneously recorded ECG to correlate it to the reconstruction. We developed a method to estimate cardiac motion directly from the measured CT raw data. In contrast to the ECG that is a global measurement of the heart's electrical excitation CMDC computes a local measure of heart motion at any scanned z-position using complementary projections. Using a series of data processing steps the synchronization information necessary for phase-correlated algorithms is extracted from the motion correlated CMDC signal. Evaluations of both computer simulations and patient measurements using multi-slice scanners have shown a high correlation between ECG and the CMDC signal. However, a significant phase shift between CMDC and ECG signal during scanning of the atria region was observed. Bot hthe ECG function and the CMDC signal were used for phase correlated reconstruction. Only minor differences in image quality between both approaches were detected, hence hte CMDC method might enhance the practicability of cardiac volume imaging in the future. CMDC-correlated reconstructions might also improve imaging of pericardial lung areas.

One of the most recent technological advancements in computed tomography (CT) is the introduction of multi-slice CT (MSCT). The state-of-the-art MSCT contains 16 detector rows and is capable of acquiring 16 projections simultaneously. In this paper, we propose a reconstruction algorithm that makes use of nontraditional reconstruction planes and convolution weighting. To minimize the impact of interpolation on slice-sensitivity-profile (SSP), conjugate samples are used for the projection interpolation. We use multiple convex planes as teh region of construction. This allows the generated weighting function to be smooth and differentiable. In addition, we make use of the fact that projections collected from a subset of detector rows are sufficient to perform a complete reconstruction. A convolution function is applied to the weighting function of each subset to minimize the impact of cone beam effects. The convolution function is chosen so that optimal balance is achieved between image artifact, slice-sensitivity-profile (SSP), and noise. Extensive phantom and clinical studies have been conducted to validate our approach. Our study indicates that compared to other row-interpolation based reconstruction algorithms, a 30% SSP improvement can be achieved with the proposed approach. In addition, image artifact suppression achieved with the proposed approach is on par or slightly better than the existing reconstruction algorithms. Extensive clinical studies have shown that the 16-slice scanner in conjugation with this algorithm produces nearly isotropic spatial resolution and allows much improved diagnostic image quality.

We have developed a model for transmission tomography that views the detected data as being Poisson-distributed photon counts. From this model, we derive an alternating minimization (AM) algorithm for the purpose of image reconstruction. This algorithm, which seeks to minimize an objective function (the I-divergence between the measured data and the estimated data), is particularly useful when high-density objects are present in soft tissue and standard image reconstruction algorithms fail. The approach incorporates inequality constraints on the pixel values and seeks to exploit known information about the high-density objects or other priors on the data. Because of the ill-posed nature of this problem, however, the noise and streaking artifacts in the images are not completely mitigated, even under the most ideal conditions, and some form of regularization is required. We describe a sieve-based approach, which constrains the image estimate to reside in a subset of the image space in which all images have been smoothed with a Gaussian kernel. The kernel is spatially varying and does not smooth across known boundaries in the image. Preliminary results show effective reduction of the noise and streak artifacts, but indicate that more work is needed to suppress edge overshoots.

We present a novel method for intra-frame image processing, which is applicable to a wide variety of medical imaging modalities, like X-ray angiography, X-ray fluoroscopy, magnetic resonance, or ultrasound. The method allows to reduce noise significantly - by about 4.5 dB and more - while preserving sharp image details. Moreover, selective amplification of image details is possible. The algorithm is based on a multi-resolution approach. Noise reduction is achieved by non-linear adaptive filtering of the individual band pass layers of the multi-resolution pyramid. The adaptivity is controlled by image gradients calculated from the next coarser layer of the multi-resolution pyramid representation, thus exploiting cross-scale dependencies. At sites with strong gradients, filtering is performed only perpendicular to the gradient, i.e. along edges or lines. The multi-resolution approach processes each detail on its appropriate scale so that also for low frequency noise small filter kernels are applied, thus limiting computational costs and allowing a real-time implementation on standard hardware. In addition, gradient norms are used to distinguish smoothly between “structure” and “noise only” areas, and to perform additional noise reduction and edge enhancement by selectively attenuating or amplifying the corresponding band pass coefficients.

Scale is a widely used notion in image analysis that evolved in the form of scale-space theory whose key idea is to represent and analyze an image at various resolutions. Recently, the notion of space-variant scale has drawn significant research interest. Previously, we introduced local morphometric scale using a spherical model whose major limitation was that it ignored orientation and anisotrophy making is suboptimal in many biomedical imaging applications where structures are inherently anisotropic and have mixed orientations. Here, we introduce a new idea of local scale, called tensor scale, which, at any image location, is the parametric representation of the largest ellipse (in 2D) or ellipsoid (in 3D) centered at that location that is contained in the same homogeneous region. Tensor scale is useful in spatially adapting neighborhood and controlling parameters in a space-variant and anisotropic fashion complying with orientation, anisotrophy, and thickness of local structures. Results of the method on several 2D images are presented and a few experiments are conducted to examine its behavior under rotation, varying pixel size, background inhomogeneity, and noise and blurring. Similarity of tensor scale images computed from multi-protocol images is studied.

We propose a new fast stereoradiographic 3D reconstruction method for the spine. User input is limited to few points passing through the spine on two radiographs and two line segments representing the end plates of the limiting vertebrae. A 3D spline that hints the positions of the vertebrae in space is then generated. We then use wavelet multi-scale analysis (WMSA) to automatically localize specific features in both lateral and frontal radiographs. The WMSA gives an elegant spectral investigation that leads to gradient generation and edge extraction. Analysis of the information contained at several scales leads to the detection of 1) two curves enclosing the vertebral bodies' walls and 2) inter-vertebral spaces along the spine. From this data, we extract four points per vertebra per view, corresponding to the corners of the vertebral bodies. These points delimit a hexahedron in space where we can match the vertebral body. This hexahedron is then passed through a 3D statistical database built using local and global information generated from a bank of normal and scoliotic spines. Finally, models of the vertebrae are positioned with respect to these landmarks, completing the 3D reconstruction.

We present a new wavelet-based strategy for autonomous feature extraction and segmentation of cardiac structures in dynamic ultrasound images. Image sequences subjected to a multidimensional (2D plus time) wavelet transform yield a large number of individual subbands, each coding for partial structural and motion information of the ultrasound sequence. We exploited this fact to create an analysis strategy for autonomous analysis of cardiac ultrasound that builds on shape- and motion specific wavelet subband filters. Subband selection was in an automatic manner based on subband statistics. Such a collection of predefined subbands corresponds to the so-called footprint of the target structure and can be used as a multidimensional multiscale filter to detect and localize the target structure in the original ultrasound sequence. Autonomous, unequivocal localization by the autonomous algorithm is then done using a peak finding algorithm, allowing to compare the findings with a reference standard. Image segmentation is then possible using standard region growing operations. To test the feasibility of this multiscale footprint algorithm, we tried to localize, enhance and segment the mitral valve autonomously in 182 non-selected clinical cardiac ultrasound sequences. Correct autonomous localization by the algorithm was feasible in 165 of 182 reconstructed ultrasound sequences, using the experienced echocardiographer as reference. This corresponds to a 91% accuracy of the proposed method in unselected clinical data. Thus, multidimensional multiscale wavelet footprints allow successful autonomous detection and segmentation of the mitral valve with good accuracy in dynamic cardiac ultrasound sequences which are otherwise difficult to analyse due to their high noise level.

Image intensity standardization is a recently developed post-processing method designed for correcting acquisition-to-acquisition signal intensity variations inherent in MR images. Inhomogeneity correction is a method used to suppress the low frequency background non-uniformities of the image domain that exist in MR images. Both these procedures have important implications for MR image analysis. The effects of these post-processing operations on improvement of image quality in isolation has been well documented [1-11]. However, the combined effects of these two processes on MR images and how the processes influence each other have not been studied thus far. In this paper, we evaluate the effect of inhomogeneity correction followed by standardization on MR images and vice-versa in order to determine the best sequence to follow for enhancing image quality. Our results indicate that improved standardization can be achieved by preceding it with inhomogeneity correction. There is no statistically significant difference in image quality obtained between the results of standardization followed by correction and that of correction followed by standardization from the perspective of inhomogeneity correction. The correction operation was found to bias the effect of standardization. We demonstrated this bias both qualitatively and quantitatively. Standardization on the other hand did not influence the correction operation. It was also found that longer sequences of repeated correction and standardization did not considerably improve image quality.

Inhomogeneity of static magnetic field, induced by object susceptibility, is unavoidable in magnetic resonance imaging (MRI). This inhomogeneity generates distortions in both image geometry and its intensity. Based on node magnetic voltage values, a fast Finite Difference Method (FDM) is developed for susceptibility-induced mapping of magnetic field inhomogeneity and applied to simulated MRI data. Its accuracy and speed of convergence are evaluated by comparing the method to Finite Elements Method (FEM), which had been validated experimentally. Effects of inhomogeneity on Spin Echo (SE) MRI are simulated using the proposed field calculation method. Also, a pixel based (direct) method as well as a grid based (indirect) method for removing the effects are developed. The fast execution of the algorithm stems from the multi-resolution nature of the proposed method. The main advantage of the proposed method is that it does not need any data except for the image itself. Efficiency of both correction methods in distortion removal is investigated.

In MRI, image intensity non-uniformity is an adverse phenomenon that increases inter-tissue overlapping. The aim of this study was to provide a novel general framework, named regularized feature condensing (RFC), for condensing the distribution of image features and apply it to correct intensity non-uniformity via spatial regularization. The proposed RCF method is an iterative procedure, which consists of four basic steps. First, creation of a feature space, which consists of multi-spectral image intensities and corresponding second derivatives. Second, estimation of the intensity condensing map in feature space, i.e. the estimation of the increase of feature probability densities by a well-established mean shift procedure. Third, regularization of intensity condensing map in image space, which yields the estimation of intensity non-uniformity. Fourth, applying the estimation of non-uniformity correction to the input image. In this way, the intensity distributions of distinct tissues are gradually condensed via spatial regularization. The method was tested on simulated and real MR brain images for which gold standard segmentations were available. The results showed that the method did not induce additional intensity variations in simulated uniform images and efficiently removed intensity non-uniformity in real MR brain images. The proposed RCF method is a powerful fully automated intensity non-uniformity correction method that makes no a prior assumptions on the image intensity distribution and provides non-parametric non-uniformity correction.

Precise density measurement of pulmonary nodules with CT is an important prerequisite if the measurement of contrastenhancement is to be used to assess if a nodule is benign or malignant. The precision of a volume-based 3D measurement method was compared to the standard 2D method currently used in clinical practice. Two consecutive low-dose CT scans (inter-scan delay a few minutes) were obtained from 10 patients with 75 pulmonary nodules (size 5 - 32 mm). A four-slice CT was used (Siemens Somatom VZ, collimation 4 x 1 mm, normalized pitch 1.75, slice thickness 1.25 mm, reconstruction interval 0.8 mm). Mean density of each nodule was determined independently from both scans with two methods: 1) an automatic 3D segmentation method; 2) the standard 2D method as proposed in the literature and currently used in clinical practice, (3 mm slice thickness, oval region of interest). ROC analysis was used to compare these methods for the detection of an enhancement of 10, 30 and 50 Hounsfield units (HU). The mean absolute measurement error (± standard deviation) was 9.9 HU (±14.4 HU) for the 3D method and 26.4 HU (±42.0 HU) for the 2D method. ROC analysis yielded A_Z values of 0.723 / 0.932 / 0.982 for the 3D method and 0.609 /0.773 / 0.850 for the 2D method for the detection of 10 / 30 / 50 HU enhancement respectively. Volume-based density determination has a significantly higher reproducibility than the currently used 2D ROI approach and should preferentially be used for enhancement measurements in pulmonary nodules.

We are developing an automated computer technique to describe calcifications in mammograms according to the BI-RADS lexicon. We evaluated this technique by its agreement with radiologists' description of the same lesions. Three expert mammographers reviewed our database of 90 cases of digitized mammograms containing clustered microcalcifications and described the calcifications according to BI-RADS. In our study, the radiologists used only 4 of the 5 calcification distribution descriptors and 5 of the 14 calcification morphology descriptors contained in BI-RADS. Our computer technique was therefore designed specifically for these 4 calcification distribution descriptors and 5 calcification morphology descriptors. For calcification distribution, 4 linear discriminant analysis (LDA) classifiers were developed using 5 computer-extracted features to produce scores of how well each descriptor describes a cluster. Similarly, for calcification morphology, 5 LDAs were designed using 10 computer-extracted features. We trained the LDAs using only the BI-RADS data reported by the first radiologist and compared the computer output to the descriptor data reported by all 3 radiologists (for the first radiologist, the leave-one-out method was used). The computer output consisted of the best calcification distribution descriptor and the best 2 calcification morphology descriptors. The results of the comparison with the data from each radiologist, respectively, were: for calcification distribution, percent agreement, 74%, 66%, and 73%, kappa value, 0.44, 0.36, and 0.46; for calcification morphology, percent agreement, 83%, 77%, and 57%, kappa value, 0.78, 0.70, and 0.44. These results indicate that the proposed computer technique can select BI-RADS descriptors in good agreement with radiologists.

PURPOSE: We compare mammographic mass classification performance between a backpropagation neural network (BNN), expert radiologists, and residents. Our goal is to reduce false negatives during routine reading of mammograms. METHODS: 160 cases from 3 different institutions were used. Each case contained at least one mass and had an accompanying biopsy result. Masses were extracted using region growing with seed locations identified by an expert radiologist. 10 texture and shape based features (area, perimeter, compactness, radial length, spiculation, mean/standard deviation of radial length, minimum/maximum axis, and boundary roughness) were used as inputs to a three-layer BNN. Shape features were computed on the boundary of the mass region; texture features were computed from the pixel values inside the mass. 140 cases were used for training the BNN and the remaining 20 cases were used for testing. The testing set was diagnosed by three expert radiologists, three residents, and the BNN. We evaluated the human readers and the BNN by computing the area under the ROC curve (AUC). RESULTS: The AUC was 0.923 for the BNN, 0.846 for the expert radiologists, and 0.648 for the residents. These results illustrate the promise of using BNN as a physician’s assistant for breast mass classification.

We have been developed a computer-aided diagnosis (CAD) system in the lung cancer detection from a low-dose single-slice CT scanner. The objective of this study is to solve three problems of the conventional CAD system; application of image obtained by other CT scanner, diagnostic procedure for the ground glass shadow less than 5 mm in diameters, and diagnostic procedure for nodule in contact with blood vessels. We analyzed characteristics between each CT images, and pattern of blood vessels. The structural analysis procedure using three-dimensional data is the newly added process. The diagnostic rules to detect nodule consist of the four classes, which are divided by size and CT value. We applied two lung cancer databases; 55 nodules of TCT-900S and 67 nodules of Asteion. The present result from the former database achieved a sensitivity of 94.5%, the latter database achieved a sensitivity of 90.0%. Most of false negative cases had two cases which are a nodule overlapped by blood vessels and a nodule on mediastinum.

A study was performed to test whether automated computer analysis of low-dose helical CT scans can accurately estimate the degree of emphysema. We characterized the severity of emphysema on low-dose high-resolution (2.5 mm slice thickness) CT scans into 4 categories (normal, mild, moderate, and severe) as determined by a thoracic radiologist. From our database we chose 80 cases (20 within each category) for analysis. Our analysis system segments the lung parenchyma from surrounding structures and computes an emphysema index as the volumetric percentage of emphysema for the entire lung volume using a dual thresholding technique. One-way analysis of variance was used to assess the emphysema index. For those cases classified as normal, the emphysema index was 11.74 ± 1.24 (mean ± sem), for mild it was 15.00 ± 1.31, for moderate it was 16.91 ± 1.72, and for severe it was 26.77 ± 1.73. The differences were statistically significant (p < 0.0001) and showed an increasing score with increasing severity of emphysema. Our system provides a useful index of the degree of emphysema present. Use of the system allows subjects undergoing lung cancer screening studies to have the extent of their emphysema quantified on a year-to-year basis.

The purpose of this study is to investigate the efficacy of image features versus likelihood features of tumor boundaries for differentiating benign and malignant tumors and to compare the effectiveness of two neural networks in the classification study: (1) circular processing-based neural network and (2) conventional Multilayer Perceptron (MLP). The segmentation method used is an adaptive region growing technique coupled with a fuzzy shadow approach and maximum likelihood analyzer. Intensity, shape, texture, and likelihood features were calculated for the extracted Region of Interest (ROI). We performed these studies: experiment number 1 utilized image features used as inputs and the MLP for classification, experiment number 2 utilized image features used as inputs and the neural net with circular processing for classification, and experiment number 3 used likelihood values as inputs and the MLP for classification. The experiments were validated using an ROC methodology. We have tested these methods on 51 mammograms using a leave-one-case-out experiment (i.e., Jackknife procedure). The A_z values for the four experiments were as follows: 0.66 in experiment number 1, 0.71 in experiment number 2, and 0.84 in experiment number 3.

In this paper, we propose a new computer aided detection (CAD) technique to utilize both global and local shape information of the colon wall for detection of colonic polyps. Firstly, the whole colon wall is extracted by our mixture-based image segmentation method. This method uses partial volume percentages to represent the distribution of different materials in each voxel, so it provides the most accurate information on the colon wall, especially the mucosa layer. Local geometrical measure of the colon mucosa layer is defined by the curvature and gradient information extracted from the segmented colon-wall mixture data. Global shape information is provided by applying an improved linear integral convolution operation to the mixture data. The CAD technique was tested on twenty patient datasets. The local geometrical measure extracted from the mixture segmentation represents more accurately the polyp variation than that extracted from conventional label classification, leading to improved detection. The added global shape information further improves the polyp detection.

We have developed a computerized method to automatically identify lung nodules in thoracic computed tomography (CT) scans. Since the ultimate goal of such a method is to improve human detection performance, the process through which computer results are conveyed to the radiologist must be considered. Detection results are presented through an interface that automatically places a circle around the detected structure in only one section in which that structure may appear. Consequently, an inappropriate choice of section could result in an actual nodule detected by the computer but not properly indicated to the radiologist, thus reducing the potential positive impact of that detection on the radiologist’s decision-making process. The automated detection method was applied to 38 diagnostic CT scans with an overall sensitivity of 71% and 0.5 false-positive detections per section; however, when these results were converted automatically to annotations on the output images for human visualization, 8.6% of the computer-detected nodules received annotations that failed to encompass a portion of the actual nodule. Thus, the "effective sensitivity" of the automated detection method (i.e., a performance paradigm that considers the eventual human interaction with system output) was reduced.

We are developing a computer-aided scheme for the detection of colonic polyps and masses in CT colonography. The colon is extracted automatically from CT images by use of a knowledge-guided technique. The detection of polyps and masses is based on shape index and curvedness features. A feature-guided segmentation technique is used to extract the regions of detected polyps. A quadratic discriminant classifier is used for reducing false-positive detections and for determining the final output based on shape index, gradient concentration, and CT value features. To evaluate the technique, we performed CT colonography for 72 patients with cleansed colons and by use of a standard technique with helical CT scanning. Thirteen patients had a total of 20 polyps measuring 5-12mm, and four patients had 4 masses measuring 25-40 mm in diameter. In a by-polyp(by-mass) leave-one-out evaluation, the CAD scheme detected 95% of the polyps(all masses) with an average of 1.5(0.5) false-positive detections per patient. These preliminary results suggest that our CAD scheme is potentially a useful tool for providing rapid interpretation and high diagnostic accuracy for CT colonography.

We have already developed a prototype of computer-aided diagnosis (CAD) system that can automatically detect suspicious shadows from Chest CT images. But the CAD system cannot detect Ground-Grass-Attenuation perfectly. In many cases, this reason depends on the inaccurate extraction of the region of interests (ROI) that CAD system analyzes, so we need to improve it. In this paper, we propose a method of an accurate extraction of the ROI, and compare proposed method to ordinary method that have used in CAD system. Proposed Method is performed by application of the three steps. Firstly we extract lung area using threshold. Secondly we remove the slowly varying bias field using flexible Opening Filter. This Opening Filter is calculated by the combination of the ordinary opening value and the distribution which CT value and contrast follow. Finally we extract Region of Interest using fuzzy clustering. When we applied proposal method to Chest CT images, we got a good result in which ordinary method cannot achieve. In this study we used the Helical CT images that are obtained under the following measurement: 10mm beam width; 20mm/sec table speed; 120kV tube voltage; 50mA tube current; 10mm reconstruction interval.

A colon polyp phantom, 28 cm long and 5 cm in diameter, was constructed by inflating a latex ultrasound transducer cover. Four round pieces of ham (3, 6, 9, 12 mm diameter) were imbedded in the outer membrane surface of the phantom and then were tied by string at the base to simulate pedunculated polyps. Three more pieces of ham (3, 6, 9 mm) were impressed and taped on the outer surface to simulate sessile polyps. The circumference of the phantom was constricted by string at four evenly spaced locations to simulate haustral folds. The phantom was placed in a water bath and was modified by infusing water into the lumen or by partially deflating the lumen, and then rescanned. CT images were obtained in a multi-slice CT (4x1 mm collimation, 0.5s scan, 120 Kvp, 90 mAs, 1 mm slice thickness). CT images were processed with our computer-aided detection program. First, the three-dimensional colonic boundary and inner structure were segmented. From this segmented region, soft-tissue structures were extracted and labeled to generate candidates. Shape features were evaluated along with geometric constraints. Three-dimensional region-growing and morphologic matching processes were applied to refine and classify the candidates. The detected polyps were compared with the true polyps in the phantom or known polyps in clinical cases to calculate the sensitivity and false positives.

We developed an ensemble classifier for the task of computer-aided diagnosis of breast microcalcification clusters,which are very challenging to characterize for radiologists and computer models alike. The purpose of this study is to help radiologists identify whether suspicious calcification clusters are benign vs. malignant, such that they may potentially recommend fewer unnecessary biopsies for actually benign lesions. The data consists of mammographic features extracted by automated image processing algorithms as well as manually interpreted by radiologists according to a standardized lexicon. We used 292 cases from a publicly available mammography database. From each cases, we extracted 22 image processing features pertaining to lesion morphology, 5 radiologist features also pertaining to morphology, and the patient age. Linear discriminant analysis (LDA) models were designed using each of the three data types. Each local model performed poorly; the best was one based upon image processing features which yielded ROC area index A_Z of 0.59 ± 0.03 and partial A_Z above 90% sensitivity of 0.08 ± 0.03. We then developed ensemble models using different combinations of those data types, and these models all improved performance compared to the local models. The final ensemble model was based upon 5 features selected by stepwise LDA from all 28 available features. This ensemble performed with A_Z of 0.69 ± 0.03 and partial A_Z of 0.21 ± 0.04, which was statistically significantly better than the model based on the image processing features alone (p<0.001 and p=0.01 for full and partial A_Z respectively). This demonstrated the value of the radiologist-extracted features as a source of information for this task. It also suggested there is potential for improved performance using this ensemble classifier approach to combine different sources of currently available data.

Lung Cancer is known as one of the most difficult cancers to cure. The detection of lung cancer in its early stage can be helpful for medical treatment to danger. However, mass screening based on helical CT images brings a considerable number of images to diagnosis, the time-consuming fact makes it difficult to be used in the clinic. To increase the efficiency of the mass screening process, we developed a Computer-aided diagnosis (CAD) system, which can detect nodules at high speed. It takes 17 seconds per case (35 images) to detect nodules. In this paper, we describe the development of this CAD system and specifications.

Mammography is an effective tool for the early detection of breast cancer; however, most women referred for biopsy based on mammographic findings do not, have cancer. This study is part of an ongoing effort to reduce the number of benign cases referred for biopsy by developing tools to aid physicians in classifying suspicious lesions. Specifically, this study examines the use of an Evolutionary Programming (EP)/Adaptive Boosting (AB) hybrid, specifically modified to focus on improving the performance of computer-assisted diagnostic (CAD) tools at high specificity levels (missing few or no cancers). An EP/AB hybrid developed by the authors and used in previous studies was modified with two new fitness functions: 1) a function which favored networks with the high PPV values at thresholds corresponding to high sensitivities and 2) a function which favored networks with the highest partial ROC Az (normalized area about 90% sensitivity). The modified hybrid with specialized fitness functions was evaluated using k-fold cross-validation against two real-word mammogram data sets. Results indicate that the number of benign cases referred for biopsy might be reduced by over a third, while missing no cancers. If sensitivity is allowed to decrease to 97% (missing 3% of the cancers), the number of spared biopsies could be raised to over half.

This work presents a classifier for mammographic masses using the wavelet transform as characteristics generator. It considers the BI-RADS classification, dividing mass according to their shapes: circulate, nodular and speculate. We developed procedures with two steps: the first involves a model applying one wavelet technique performing the contours analysis with simulated mass images. This procedure was used to choose the best wavelet that could generate the desired characteristics. The second procedure had the objective of applying the chosen wavelet to masses from segmented images. Both methods have as answers three classes of shape. A root-mean-square function is applied to obtain the energy measure for each level of wavelet decomposition. Thus the shape feature vectors are formed with the coefficients of the details and coefficients of approximation extracted by the energy of wavelet decomposition levels. Linear Discriminan Analysis (LDA) by using Fischer Discriminant was used to reduce the number of characteristics for the feature vector. The Mahalanobis distance was used by the classifier to verify aimed the pertinence of the images for each one the previously given classes. To test actual images, the leave-one-out method was used to the classifier training. The classifier has registered good results, compared to others reports in the corresponding literature.

A novel technique for assessing local and global differences between mammographic images was developed. This method uses correlations between abstract features extracted from corresponding views to compare image properties without resorting to processes that depend on exact geometrical congruence, such as image subtraction, which have a tendency to produce excessive artifact. The method begins by normalizing both digitized mammograms, after which a series of global and local feature filters are applied to each image. Each filter calculates values characterizing a particular property of the given image, and these values, for each property of interest are arranged in a feature vector. Corresponding elements in the two feature vectors are combined to produce a difference vector that indicates the change in the particular properties between images. Features are selected which are expected to be relatively invariant with respect to breast compression.

Full-field digital mammography (FFDM) as a new breast imaging modality has potential to detect more breast cancers or to detect them at smaller sizes and earlier stages compared with screening film mammography (SFM). However, its performance needs verification, and it would pose new problems for the development of CAD methods for breast cancer detection and diagnosis. Performance evaluation of CAD systems on FFDM and SFM has been conducted in this study, respectively. First, an adaptive CAD system employing a series of advanced modules has been developed on FFDM. Second, a standardization approach has been developed to make the CAD system independent of characteristics of digitizer or imaging modalities for mammography. CAD systems developed previously for SFM and developed in this study for FFDM have been evaluated on FFDM and SFM images without and with standardization, respectively, to examine the performance improvement of the CAD system developed in this study. Computerized free-response receiver operating characteristic (FROC) analysis has been adopted as performance evaluation method. Compared with previous one, the CAD system developed in this study demonstrated significantly performance improvements. However, the comparison results have shown that the performances of final CAD system in this study are not significantly different on FFDM and on SFM after standardization. It needs further study on the assessment of CAD system performance on FFDM and SFM modalities.