A comparison of neural network and fuzzy clustering techniques in segmenting magnetic resonance images of the brain

Multispectral analysis of bone lesions in the hands of patients with rheumatoid arthritis

Quantitative measures of rheumatoid arthritis (RA) disease progression can provide valuable tools for evaluation of new treatments during clinical trials. In this study, a novel multispectral (MS) MRI analysis method is presented to quantify changes in bone lesion volume (DeltaBLV) in the hands of RA patients. Image registration and MS analysis were employed to identify MS tissue class transitions between two serial MRI exams. DeltaBLV was determined from MS class transitions between two time points. The following three classifiers were investigated: (a) multivariate Gaussian (MVG), (b) k-nearest neighbor (k-NN), and (c) K-means (KM). Unlike supervised classifiers (MVG, k-NN), KM, an unsupervised classifier, does not require labeled training data, resulting in potentially greater clinical utility. All MS estimates of DeltaBLV were linearly correlated (r(p)) with manual estimates. KM and k-NN estimates also exhibited a significant rank-order correlation (r(s)) with manual estimates. For KM, r(p) = 0.94 p < 0.0001, r(s) = 0.76 p = 0.002; for k-NN, r(p) = 0.86 p = 0.0001, r(s) = 0.69 p = 0.009; and for MVG, r(p) = 0.84 p = 0.0003, r(s) = 0.49 p = 0.09. Temporal classification rates were as follows: for KM, 90.1%; for MVG, 89.5%; and for k-NN, 86.7%. KM matched the performance of k-NN, offering strong potential for use in multicenter clinical trials. This study demonstrates that MS tissue class transitions provide a quantitative measure of DeltaBLV.

Analysis of the Weighting Exponent in the FCM

The fuzzy c-means (FCM) algorithm is one of the most frequently used clustering algorithms. The weighting exponent m is a parameter that greatly influences the performance of the FCM. But there has been no theoretical basis for selecting the proper weighting exponent in the literature. In this paper, we develop a new theoretical approach to selecting the weighting exponent in the FCM. Based on this approach, we reveal the relation between the stability of the fixed points of the FCM and the data set itself. This relation provides the theoretical basis for selecting the weighting exponent in the FCM. The numerical experiments verify the effectiveness of our theoretical conclusion.

Generalized Kohonen’s competitive learning algorithms for ophthalmological MR image segmentation

Kohonen's self-organizing map is a two-layer feedforward competitive learning network. It has been used as a competitive learning clustering algorithm. In this paper, we generalize Kohonen's competitive learning (KCL) algorithm with fuzzy and fuzzy-soft types called fuzzy KCL (FKCL) and fuzzy-soft KCL (FSKCL). These generalized KCL algorithms fuse the competitive learning with soft competition and fuzzy c-means (FCM) membership functions. We then apply these generalized KCLs to MRI and MRA ophthalmological segmentations. These KCL-based MRI segmentation techniques are useful in reducing medical image noise effects using a learning mechanism. They may be particularly helpful in clinical diagnosis. Two real cases with MR image data recommended by an ophthalmologist are examined. First case is a patient with Retinoblastoma in her left eye, an inborn malignant neoplasm of the retina frequently metastasis beyond the lacrimal cribrosa. The second case is a patient with complete left side oculomotor palsy immediately after a motor vehicle accident. Her brain MRI with MRA, skull routine, orbital CT, and cerebral angiography did not reveal brainstem lesions, skull fractures, or vascular anomalies. These generalized KCL algorithms were used in segmenting the ophthalmological MRIs. KCL, FKCL and FSKCL comparisons are made. Overall, the FSKCL algorithm is recommended for use in MR image segmentation as an aid to small lesion diagnosis.

An adaptive spatial fuzzy clustering algorithm for 3-D MR image segmentation

An adaptive spatial fuzzy c-means clustering algorithm is presented in this paper for the segmentation of three-dimensional (3-D) magnetic resonance (MR) images. The input images may be corrupted by noise and intensity nonuniformity (INU) artifact. The proposed algorithm takes into account the spatial continuity constraints by using a dissimilarity index that allows spatial interactions between image voxels. The local spatial continuity constraint reduces the noise effect and the classification ambiguity. The INU artifact is formulated as a multiplicative bias field affecting the true MR imaging signal. By modeling the log bias field as a stack of smoothing B-spline surfaces, with continuity enforced across slices, the computation of the 3-D bias field reduces to that of finding the B-spline coefficients, which can be obtained using a computationally efficient two-stage algorithm. The efficacy of the proposed algorithm is demonstrated by extensive segmentation experiments using both simulated and real MR images and by comparison with other published algorithms.

Multicontext fuzzy clustering for separation of brain tissues in magnetic resonance images

A local image model is proposed to eliminate the adverse impact of both artificial and inherent intensity inhomogeneities in magnetic resonance imaging on intensity-based image segmentation methods. The estimation and correction procedures for intensity inhomogeneities are no longer indispensable because the highly convoluted spatial distribution of different tissues in the brain is taken into consideration. On the basis of the local image model, multicontext fuzzy clustering (MCFC) is proposed for classifying 2D and 3D MR data into tissues of white matter, gray matter, and cerebral spinal fluid automatically. In MCFC, multiple clustering contexts are generated for each pixel, and fuzzy clustering is independently performed in each context to calculate the degree of membership of a pixel to each tissue class. To maintain the statistical reliability and spatial continuity of membership distributions, a fusion strategy is adopted to integrate the clustering outcomes from different contexts. The fusion result is taken as the final membership value of the pixel. Experimental results on both real MR images and simulated volumetric MR data show that MCFC outperforms the classic fuzzy c-means (FCM) as well as other segmentation methods that deal with intensity inhomogeneities.

A genetic algorithm based nearest neighbor classification to breast cancer diagnosis

This paper presents an application of a hybrid approach (the genetic algorithms and the k-nearest neighbour) proposed by Ishbuchi to Wisconsin breast cancer data. For the diagnosis of breast cancer, the determination of the presence of benign/malignant breast tumors represents a very complex problem (even for an experienced cytologist). Therefore the automatic classification of benign and malignant symptoms is highly desirable as a valuable aid to assist oncologists in the decision making of the diagnosis of breast cancer. In this paper, the genetic algorithm based k-nearest neighbour method for classification of benign and malignant breast tumors is presented. The genetic-algorithm (GA) is used for finding a compact reference set by selecting a small number of reference patterns from a large number of training patterns in nearest neighbor classification. The GA simultaneously performs feature selection and pattern selection and prunes unnecessary features. The goal is to maximize the classification performance of the reference set and minimize the number of selected patterns and features. Results are also compared with a fuzzy-genetic approach where each reference patten represents a fuzzy if-then rule with a circular-cone-type membership function.

Feature-based, automated segmentation of cerebral infarct patterns using T2- and diffusion-weighted imaging

Diffusion-weighted imaging enables the diagnosis of cerebral ischemias very early, thus supporting therapies such as thrombolysis. However, morphology and tissue-characterizing parameters (e.g. relaxation times or water diffusion) may vary strongly in ischemic regions, indicating different underlying pathologic processes. As the determination of the parameters by a supervised segmentation is very time consuming, we evaluated whether different infarct patterns may be segmented by an automated, multidimensional feature-based method using a unified segmentation procedure. Ischemias were classified into 5 characteristic patterns. For each class, a 3D histogram based on T(2)- and diffusion-weighted images as well as calculated apparent diffusion coefficients (ADC) was generated from a representative data set. Healthy and pathologic tissue classes were segmented in the histogram as separate, local density maxima with freely shaped borders. Segmentation control parameters were optimized in a 3-step procedure. The method was evaluated using synthetic images as well as results of a supervised segmentation. For the analysis of cerebral ischemias, the optimal control parameter set led to sensitivities and specificities between 1.0 and 0.9.

Segmentation of MR images for computer-assisted surgery of the lumbar spine

This paper describes a segmentation algorithm designed to separate bone from soft tissue in magnetic resonance (MR) images developed for computer-assisted surgery of the spine. The algorithm was applied to MR images of the spine of healthy volunteers. Registration experiments were carried out on a physical model of a spine generated from computed tomography (CT) data of a surgical patient. Segmented CT, manually segmented MR and MR images segmented using the developed algorithm were compared. The algorithm performed well at segmenting bone from soft tissue on images taken of healthy volunteers. Registration experiments showed similar results between the CT and MR data. The MR data, which were manually segmented, performed worse on visual verification experiments than both the CT and semi-automatic segmented data. The algorithm developed performs well at segmenting bone from soft tissue in MR images of the spine as measured using registration experiments.

A modified fuzzy C-means algorithm for bias field estimation and segmentation of MRI data

In this paper, we present a novel algorithm for fuzzy segmentation of magnetic resonance imaging (MRI) data and estimation of intensity inhomogeneities using fuzzy logic. MRI intensity inhomogeneities can be attributed to imperfections in the radio-frequency coils or to problems associated with the acquisition sequences. The result is a slowly varying shading artifact over the image that can produce errors with conventional intensity-based classification. Our algorithm is formulated by modifying the objective function of the standard fuzzy c-means (FCM) algorithm to compensate for such inhomogeneities and to allow the labeling of a pixel (voxel) to be influenced by the labels in its immediate neighborhood. The neighborhood effect acts as a regularizer and biases the solution toward piecewise-homogeneous labelings. Such a regularization is useful in segmenting scans corrupted by salt and pepper noise. Experimental results on both synthetic images and MR data are given to demonstrate the effectiveness and efficiency of the proposed algorithm.

Current methods in medical image segmentation

Image segmentation plays a crucial role in many medical-imaging applications, by automating or facilitating the delineation of anatomical structures and other regions of interest. We present a critical appraisal of the current status of semi-automated and automated methods for the segmentation of anatomical medical images. Terminology and important issues in image segmentation are first presented. Current segmentation approaches are then reviewed with an emphasis on the advantages and disadvantages of these methods for medical imaging applications. We conclude with a discussion on the future of image segmentation methods in biomedical research.

Magnetic resonance image analysis by information theoretic criteria and stochastic site models

Quantitative analysis of magnetic resonance (MR) images is a powerful tool for image-guided diagnosis, monitoring, and intervention. The major tasks involve tissue quantification and image segmentation where both the pixel and context images are considered. To extract clinically useful information from images that might be lacking in prior knowledge, we introduce an unsupervised tissue characterization algorithm that is both statistically principled and patient specific. The method uses adaptive standard finite normal mixture and inhomogeneous Markov random field models, whose parameters are estimated using expectation-maximization and relaxation labeling algorithms under information theoretic criteria. We demonstrate the successful applications of the approach with synthetic data sets and then with real MR brain images.

Automatic segmentation of non-enhancing brain tumors in magnetic resonance images

Tumor segmentation from magnetic resonance (MR) images may aid in tumor treatment by tracking the progress of tumor growth and/or shrinkage. In this paper we present the first automatic segmentation method which separates non-enhancing brain tumors from healthy tissues in MR images to aid in the task of tracking tumor size over time. The MR feature images used for the segmentation consist of three weighted images (T1, T2 and proton density (PD)) for each axial slice through the head. An initial segmentation is computed using an unsupervised fuzzy clustering algorithm. Then, integrated domain knowledge and image processing techniques contribute to the final tumor segmentation. They are applied under the control of a knowledge-based system. The system knowledge was acquired by training on two patient volumes (14 images). Testing has shown successful tumor segmentations on four patient volumes (31 images). Our results show that we detected all six non-enhancing brain tumors, located tumor tissue in 35 of the 36 ground truth (radiologist labeled) slices containing tumor and successfully separated tumor regions from physically connected CSF regions in all the nine slices. Quantitative measurements are promising as correspondence ratios between ground truth and segmented tumor regions ranged between 0.368 and 0.871 per volume, with percent match ranging between 0.530 and 0.909 per volume.

Multispectral analysis of the temporal evolution of cerebral ischemia in the rat brain

A major difficulty in staging and predicting ischemic brain injury by magnetic resonance (MR) imaging is the time-varying nature of the MR parameters within the ischemic lesion. A new multispectral (MS) approach is described to characterize cerebral ischemia in a time-independent fashion. MS analysis of five MR parameters (mean diffusivity, diffusion anisotropy, T2, proton density, and perfusion) was employed to characterize the progression of ischemic lesion in the rat brain following 60 minutes of transient focal ischemia. k-Means (KM) and fuzzy c-means (FCM) classification methods were employed to define the acute and subacute ischemic lesion. KM produced an estimate of lesion volume that was highly correlated with postmortem infarct volume, independent of the age of the lesion. Overall classification rates for KM exceeded FCM at acute and subacute time points as follows: KM, 90.5%, 94.4%, and 95. 9%; FCM, 82.4%, 90.6%, and 82.6% (for 45 minutes, 180 minutes, and 24-120 hours post MCAO groups). MS analysis also offers a formal method of combining diffusion and perfusion parameters to provide an estimate of the ischemic penumbra (KM classification rate = 70.3%). J. Magn. Reson. Imaging 2000;12:842-858.

Multi-level adaptive segmentation of multi-parameter MR brain images

MR brain image segmentation into several tissue classes is of significant interest to visualize and quantify individual anatomical structures. Traditionally, the segmentation is performed manually in a clinical environment that is operator dependent and may be difficult to reproduce. Though several algorithms have been investigated in the literature for computerized automatic segmentation of MR brain images, they are usually targeted to classify image into a limited number of classes such as white matter, gray matter, cerebrospinal fluid and specific lesions. We present a novel model-based method for the automatic segmentation and classification of multi-parameter MR brain images into a larger number of tissue classes of interest. Our model employs 15 brain tissue classes instead of the commonly used set of four classes, which were of clinical interest to neuroradiologists for following-up with patients suffering from cerebrovascular deficiency (CVD) and/or stroke. The model approximates the spatial distribution of tissue classes by a Gauss Markov random field and uses the maximum likelihood method to estimate the class probabilities and transitional probabilities for each pixel of the image. Multi-parameter MR brain images with T(1), T(2), proton density, Gd+T(1), and perfusion imaging were used in segmentation and classification. In the development of the segmentation model, true class-membership of measured parameters was determined from manual segmentation of a set of normal and pathologic brain images by a team of neuroradiologists. The manual segmentation was performed using a human-computer interface specifically designed for pixel-by-pixel segmentation of brain images. The registration of corresponding images from different brains was accomplished using an elastic transformation. The presented segmentation method uses the multi-parameter model in adaptive segmentation of brain images on a pixel-by-pixel basis. The method was evaluated on a set of multi-parameter MR brain images of a twelve-year old patient 48h after suffering a stroke. The results of classification as compared to the manual segmentation of the same data show the efficacy and accuracy of the presented methods as well as its capability to create and learn new tissue classes.

A multiresolution image segmentation technique based on pyramidal segmentation and fuzzy clustering

In this paper, an unsupervised image segmentation technique is presented, which combines pyramidal image segmentation with the fuzzy c-means clustering algorithm. Each layer of the pyramid is split into a number of regions by a root labeling technique, and then fuzzy c-means is used to merge the regions of the layer with the highest image resolution. A cluster validity functional is used to find the optimal number of objects automatically. Segmentation of a number of synthetic as well as clinical images is illustrated and two fully automatic segmentation approaches are evaluated, which determine the left ventricular volume (LV) in 140 cardiovascular magnetic resonance (MR) images. First fuzzy c-means is applied without pyramids. In the second approach the regions generated by pyramidal segmentation are merged by fuzzy c-means. The correlation coefficients of manually and automatically defined LV lumen of all 140 and 20 end-diastolic images were equal to 0.86 and 0.79, respectively, when images were segmented with fuzzy c-means alone. These coefficients increased to 0.90 and 0.93 when the pyramidal segmentation was combined with fuzzy c-means. This method can be applied to any dimensional representation and at any resolution level of an image series. The evaluation study shows good performance in detecting LV lumen in MR images.

Adaptive fuzzy segmentation of magnetic resonance images

An algorithm is presented for the fuzzy segmentation of two-dimensional (2-D) and three-dimensional (3-D) multispectral magnetic resonance (MR) images that have been corrupted by intensity inhomogeneities, also known as shading artifacts. The algorithm is an extension of the 2-D adaptive fuzzy C-means algorithm (2-D AFCM) presented in previous work by the authors. This algorithm models the intensity inhomogeneities as a gain field that causes image intensities to smoothly and slowly vary through the image space. It iteratively adapts to the intensity inhomogeneities and is completely automated. In this paper, we fully generalize 2-D AFCM to three-dimensional (3-D) multispectral images. Because of the potential size of 3-D image data, we also describe a new faster multigrid-based algorithm for its implementation. We show, using simulated MR data, that 3-D AFCM yields lower error rates than both the standard fuzzy C-means (FCM) algorithm and two other competing methods, when segmenting corrupted images. Its efficacy is further demonstrated using real 3-D scalar and multispectral MR brain images.

Abdominal organ segmentation using texture transforms and a Hopfield neural network

Abdominal organ segmentation is highly desirable but difficult, due to large differences between patients and to overlapping grey-scale values of the various tissue types. The first step in automating this process is to cluster together the pixels within each organ or tissue type. We propose to form images based on second-order statistical texture transforms (Haralick transforms) of a CT or MRI scan. The original scan plus the suite of texture transforms are then input into a Hopfield neural network (HNN). The network is constructed to solve an optimization problem, where the best solution is the minima of a Lyapunov energy function. On a sample abdominal CT scan, this process successfully clustered 79-100% of the pixels of seven abdominal organs. It is envisioned that this is the first step to automate segmentation. Active contouring (e.g., SNAKE's) or a back-propagation neural network can then be used to assign names to the clusters and fill in the incorrectly clustered pixels.

Feature extraction for MRI segmentation

Magnetic resonance images (MRIs) of the brain are segmented to measure the efficacy of treatment strategies for brain tumors. To date, no reproducible technique for measuring tumor size is available to the clinician, which hampers progress of the search for good treatment protocols. Many segmentation techniques have been proposed, but the representation (features) of the MRI data has received little attention. A genetic algorithm (GA) search was used to discover a feature set from multi-spectral MRI data. Segmentations were performed using the fuzzy c-means (FCM) clustering technique. Seventeen MRI data sets from five patients were evaluated. The GA feature set produces a more accurate segmentation. The GA fitness function that achieves the best results is the Wilks's lambda statistic when applied to FCM clusters. Compared to linear discriminant analysis, which requires class labels, the same or better accuracy is obtained by the features constructed from a GA search without class labels, allowing fully operator independent segmentation. The GA approach therefore provides a better starting point for the measurement of the response of a brain tumor to treatment.

Anatomical model matching with fuzzy implicit surfaces for segmentation of thoracic volume scans

Many segmentation methods for thoracic volume data require manual input in the form of a seed point, initial contour, volume of interest etc. The aim of the work presented here is to further automate this segmentation initialization step. In this paper an anatomical modeling and matching method is proposed to coarsely segment thoracic volume data into anatomically labeled regions. An anatomical model of the thorax is constructed in two steps: 1) individual organs are modeled with blended fuzzy implicit surfaces and 2) the single organ models are grouped into a tree structure with a solid modeling technique named constructive solid geometry (CSG). The combination of CSG with fuzzy implicit surfaces allows a hierarchical scene description by means of a boundary model, which characterizes the scene volume as a boundary potential function. From this boundary potential, an energy function is defined which is minimal when the model is registered to the tissue-air transitions in thoracic magnetic resonance imaging (MRI) data. This allows automatic registration in three steps: feature detection, initial positioning and energy minimization. The model matching has been validated in phantom simulations and on 15 clinical thoracic volume scans from different subjects. In 13 of these sets the matching method accurately partitioned the image volumes into a set of volumes of interest for the heart, lungs, cardiac ventricles, and thorax outlines. The method is applicable to segmentation of various types of thoracic MR-images, provided that a large part of the thorax is contained in the image volume.

Segmentation of magnetic resonance images using fuzzy algorithms for learning vector quantization

This paper evaluates a segmentation technique for magnetic resonance (MR) images of the brain based on fuzzy algorithms for learning vector quantization (FALVQ). These algorithms perform vector quantization by updating all prototypes of a competitive network through an unsupervised learning process. Segmentation of MR images is formulated as an unsupervised vector quantization process, where the local values of different relaxation parameters form the feature vectors which are represented by a relatively small set of prototypes. The experiments evaluate a variety of FALVQ algorithms in terms of their ability to identify different tissues and discriminate between normal tissues and abnormalities.

Determination of focal ischemic lesion volume in the rat brain using multispectral analysis

Multispectral (MS) analysis was used to determine the ischemic lesion volume in the rat brain after permanent middle cerebral arterial occlusion. MS analysis used a four-dimensional MS model consisting of an estimate of the average apparent diffusion coefficient of water (ADC(av)), T2, proton density, and perfusion. Four classification methods were investigated: (a) multivariate gaussian (MVG); (b) k-nearest neighbor (k-NN); (c) k-means (KM); and (d) fuzzy c-means (FCM). MVG and k-NN classifiers are supervised methods requiring labeled training data to characterize the stroke lesion. Unsupervised classifiers (KM, FCM) do not require previous statistics or labeled training data, resulting in potentially greater clinical usefulness. All MS methods provided significant correlation with postmortem findings beyond the use of ADC(av) alone (partial correlation given the ADC(av) estimate: MVG, .66; k-NN, .75; KM, .68; FCM, .70). This study demonstrates that MS analysis provides an improved estimate of ischemic lesion volume over that obtained from ADC alone.

A fast and accurate method for registration of MR images of the head

This paper proposes a new fully automated technique that can be used for the registration of medical images of the head. The method uses Chebyshev polynomials in order to approximate and then minimize a novel multiresolutional, signal intensity independent disparity function, which can generally be defined as the mean squared value of the mean weighted ratio of two images. This function is explicitly computed for n Chebyshev points in a geometric transformation parameter interval [-A, +A] transformation units and is approximated using the Chebyshev polynomials for all other points in the interval. For 3D T2-T1 weighted MR registration, 120 experiments with studies from ten patients were performed and showed that n = 4 Chebyshev points for A = 18 transformation units give mean rotational error 0.36 degrees and a mean translational error 0.36 mm. The different noise conditions did not affect the performance of the method. We conclude that the method is suitable for routine clinical applications and that it has significant potential for future development and improvement.

Quantification and Segmentation of Brain Tissues from MR Images: A Probabilistic Neural Network Approach

This paper presents a probabilistic neural network based technique for unsupervised quantification and segmentation of brain tissues from magnetic resonance images. It is shown that this problem can be solved by distribution learning and relaxation labeling, resulting in an efficient method that may be particularly useful in quantifying and segmenting abnormal brain tissues where the number of tissue types is unknown and the distributions of tissue types heavily overlap. The new technique uses suitable statistical models for both the pixel and context images and formulates the problem in terms of model-histogram fitting and global consistency labeling. The quantification is achieved by probabilistic self-organizing mixtures and the segmentation by a probabilistic constraint relaxation network. The experimental results show the efficient and robust performance of the new algorithm and that it outperforms the conventional classification based approaches.

Robust parameter estimation of intensity distributions for brain magnetic resonance images

This paper presents two new methods for robust parameter estimation of mixtures in the context of magnetic resonance (MR) data segmentation. The head is constituted of different types of tissue that can be modeled by a finite mixture of multivariate Gaussian distributions. Our goal is to estimate accurately the statistics of desired tissues in presence of other ones of lesser interest. These latter can be considered as outliers and can severely bias the estimates of the former. For this purpose, we introduce a first method, which is an extension of the expectation-maximization (EM) algorithm, that estimates parameters of Gaussian mixtures but incorporates an outlier rejection scheme which allows to compute the properties of the desired tissues in presence of atypical data. The second method is based on genetic algorithms and is well suited for estimating the parameters of mixtures of different kind of distributions. We use this property by adding a uniform distribution to the Gaussian mixture for modeling the outliers. The proposed genetic algorithm can efficiently estimate the parameters of this extended mixture for various initial settings. Also, by changing the minimization criterion, estimates of the parameters can be obtained by histogram fitting which considerably reduces the computational cost. Experiments on synthetic and real MR data show that accurate estimates of the gray and white matters parameters are computed.

Automatic tumor segmentation using knowledge-based techniques

A system that automatically segments and labels glioblastoma-multiforme tumors in magnetic resonance images (MRI's) of the human brain is presented. The MRI's consist of T1-weighted, proton density, and T2-weighted feature images and are processed by a system which integrates knowledge-based (KB) techniques with multispectral analysis. Initial segmentation is performed by an unsupervised clustering algorithm. The segmented image, along with cluster centers for each class are provided to a rule-based expert system which extracts the intracranial region. Multispectral histogram analysis separates suspected tumor from the rest of the intracranial region, with region analysis used in performing the final tumor labeling. This system has been trained on three volume data sets and tested on thirteen unseen volume data sets acquired from a single MRI system. The KB tumor segmentation was compared with supervised, radiologist-labeled "ground truth" tumor volumes and supervised k-nearest neighbors tumor segmentations. The results of this system generally correspond well to ground truth, both on a per slice basis and more importantly in tracking total tumor volume during treatment over time.

Precise segmentation of the lateral ventricles and caudate nucleus in MR brain images using anatomically driven histograms

This paper demonstrates a time-saving, automated method that helps to segment the lateral ventricles and caudate nucleus in T1-weighted coronal magnetic resonance (MR) brain images of normal control subjects. The method involves choosing intensity thresholds by using anatomical information and by locating peaks in histograms. To validate the method, the lateral ventricles and caudate nucleus were segmented in three brain scans by four experts, first using an established method involving isointensity contours and manual editing, and second using automatically generated intensity thresholds as an aid to the established method. The results demonstrate both time savings and increased reliability.

MRI measurement of brain tumor response: comparison of visual metric and automatic segmentation

An automatic magnetic resonance imaging (MRI) multispectral segmentation method and a visual metric are compared for their effectiveness to measure tumor response to therapy. Automatic response measurements are important for multicenter clinical trials. A visual metric such as the product of the largest diameter and the largest perpendicular diameter of the tumor is a standard approach, and is currently used in the Radiation Treatment Oncology Group (RTOG) and the Eastern Cooperative Oncology Group (EGOG) clinical trials. In the standard approach, the tumor response is based on the percentage change in the visual metric and is categorized into cure, partial response, stable disease, or progression. Both visual and automatic methods are applied to six brain tumor cases (gliomas) of varying levels of segmentation difficulty. The analyzed data were serial multispectral MR images, collected using MR contrast enhancement. A fully automatic knowledge guided method (KG) was applied to the MRI multispectral data, while the visual metric was taken from the MRI films using the T1 gadolinium enhanced image, with repeat measurements done by two radiologists and two residents. Tumor measurements from both visual and automatic methods are compared to "ground truth," (GT) i.e., manually segmented tumor. The KG method was found to slightly overestimate tumor volume, but in a consistent manner, and the estimated tumor response compared very well to hand-drawn ground truth with a correlation coefficient of 0.96. In contrast, the visually estimated metric had a large variation between observers, particularly for difficult cases, where the tumor margins are not well delineated. The inter-observer variation for the measurement of the visual metric was only 16%, i.e., observers generally agreed on the lengths of the diameters. However, in 30% of the studied cases no consensus was found for the categorical tumor response measurement, indicating that the categories are very sensitive to variations in the diameter measurements. Moreover, the method failed to correctly identify the response in half of the cases. The data demonstrate that automatic 3D methods are clearly necessary for objective and clinically meaningful assessment of tumor volume in single or multicenter clinical trials.

Statistical approach to segmentation of single-channel cerebral MR images

A statistical model is presented that represents the distributions of major tissue classes in single-channel magnetic resonance (MR) cerebral images. Using the model, cerebral images are segmented into gray matter, white matter, and cerebrospinal fluid (CSF). The model accounts for random noise, magnetic field inhomogeneities, and biological variations of the tissues. Intensity measurements are modeled by a finite Gaussian mixture. Smoothness and piecewise contiguous nature of the tissue regions are modeled by a three-dimensional (3-D) Markov random field (MRF). A segmentation algorithm, based on the statistical model, approximately finds the maximum a posteriori (MAP) estimation of the segmentation and estimates the model parameters from the image data. The proposed scheme for segmentation is based on the iterative conditional modes (ICM) algorithm in which measurement model parameters are estimated using local information at each site, and the prior model parameters are estimated using the segmentation after each cycle of iterations. Application of the algorithm to a sample of clinical MR brain scans, comparisons of the algorithm with other statistical methods, and a validation study with a phantom are presented. The algorithm constitutes a significant step toward a complete data driven unsupervised approach to segmentation of MR images in the presence of the random noise and intensity inhomogeneities.

Improved intracranial lesion characterization by tissue segmentation based on a 3D feature map

Our aim was to develop an accurate multispectral tissue segmentation method based on 3D feature maps. We utilized proton density (PD), T2-weighted fast spin-echo (FSE), and T1-weighted spin-echo images as inputs for segmentation. Phantom constructs, cadaver brains, an animal brain tumor model and both normal human brains and those from patients with either multiple sclerosis (MS) or primary brain tumors were analyzed with this technique. Initially, misregistration, RF inhomogeneity and image noise problems were addressed. Next, a qualified observer identified samples representing the tissues of interest. Finally, k-nearest neighbor algorithm (k-NN) was utilized to create a stack of color-coded segmented images. The inclusion of T1 based images, as a third input, produced significant improvement in the delineation of tissues. In MS, our 3D technique was found to be far superior to that based on any combination of 2D feature maps (P < 0.001). We identified at least two distinctly different classes of lesions within the same MS plaque, representing different stages of the disease process. Further, we obtained the regional distribution of MS lesion burden and followed its changes over time. Neuropsychological aberrations were the clinical counterpart of the structural changes detected in segmentation. We could also delineate the margins of benign brain tumors. In malignant tumors, up to four abnormal tissues were identified: 1) a solid tumor core, 2) a cystic component, 3) edema in the white matter, and 4) areas of necrosis and hemorrhage. Subsequent neurosurgical exploration confirmed the distribution of tissues as predicted by this analysis.

Using neural networks to automatically detect brain tumours in MR images

Computer vision has been applied to many medical imaging problems with the aim of providing clinical tools to aid medical professionals. We present work being carried out to develop one such system to automatically detect a specific type of brain tumour from head MR images. The tumour under consideration is an acoustic neuroma, which is a benign tumour occurring in the acoustic canals. The hybrid system developed integrates neural networks with more conventional techniques used for computer vision tasks. A database of MR images from 50 patients has been assembled and the acoustic neuromas present in the images have been labelled by hand. Using this data, neural networks (MLPs) have been developed to classify the images at the pixel level to achieve a targeted segmentation. The data used to train and test the MLPs developed, consists of the grey levels of a square of pixels, the pixel to be classified being the centre pixel, together with its global position in the image. The initial pixel level segmentation is refined by a series of conventional techniques. It is combined with an edge-region based segmentation and a morphological operation is applied to the result. This processing produces clusters of adjacent regions, which are considered to be candidate tumour regions. For each possible combination of these regions, features are measured and presented to neural networks which have been trained to identify structures corresponding to acoustic neuromas. Using this approach, all the acoustic neuromas are identified together with three false positive errors.

Monitoring brain tumor response to therapy using MRI segmentation

The performance evaluation of a semi-supervised fuzzy c-means (SFCM) clustering method for monitoring brain tumor volume changes during the course of routine clinical radiation-therapeutic and chemo-therapeutic regimens is presented. The tumor volume determined using the SFCM method was compared with the volume estimates obtained using three other methods: (a) a k nearest neighbor (kNN) classifier, b) a grey level thresholding and seed growing (ISG-SG) method and c) a manual pixel labeling (GT) method for ground truth estimation. The SFCM and kNN methods are applied to the multispectral, contrast enhanced T1, proton density, and T2 weighted, magnetic resonance images (MRI) whereas the ISG-SG and GT methods are applied only to the contrast enhanced T1 weighted image. Estimations of tumor volume were made on eight patient cases with follow-up MRI scans performed over a 32 week interval during treatment. The tumor cases studied include one meningioma, two brain metastases and five gliomas. Comparisons with manually labeled ground truth estimations showed that there is a limited agreement between the segmentation methods for absolute tumor volume measurements when using images of patients after treatment. The average intraobserver reproducibility for the SFCM, kNN and ISG-SG methods was found to be 5.8%, 6.6% and 8.9%, respectively. The average of the interobserver reproducibility of these methods was found to be 5.5%, 6.5% and 11.4%, respectively. For the measurement of relative change of tumor volume as required for the response assessment, the multi-spectral methods kNN and SFCM are therefore preferred over the seedgrowing method.

Watershed-based segmentation of 3D MR data for volume quantization

The aim of this work is the development of a semiautomatic segmentation technique for efficient and accurate volume quantization of Magnetic Resonance (MR) data. The proposed technique uses a 3D variant of Vincent and Soilles immersion-based watershed algorithm that is applied to the gradient magnitude of the MR data and that produces small volume primitives. The known drawback of the watershed algorithm, oversegmentation, is strongly reduced by a priori application of a 3D adaptive anisotropic diffusion filter to the MR data. Furthermore, oversegmentation is a posteriori reduced by properly merging small volume primitives that have similar gray level distributions. The outcome of the proceeding image processing steps is presented to the user for manual segmentation. Through selection of volume primitives, the user quickly segments of first slice, which contains the object of interest. Afterwards, the subsequent slices are automatically segmented by extrapolation. Segmentation results are contingently manually corrected. The proposed segmentation technique is tested on phantom objects, where segmentation errors less than 2% are observed. In addition, the technique is demonstrated on 3D MR data of the mouse head from which the cerebellum is extracted. Volumes of the mouse cerebellum and the mouse brains in toto are calculated.

Normal brain volume measurements using multispectral MRI segmentation

The performance of a supervised k-nearest neighbor (kNN) classifier and a semisupervised fuzzy c-means (SFCM) clustering segmentation method are evaluated for reproducible measurement of the volumes of normal brain tissues and cerebrospinal fluid. The stability of the two segmentation methods is evaluated for (a) operator selection of training data, (b) reproducibility during repeat imaging sessions to determine any variations in the sensor performance over time, (c) variations in the measured volumes between different subjects, and (d) variability with different imaging parameters. The variations were found to be dependent on the type of measured tissue and the operator performing the segmentations. The variability during repeat imaging sessions for the SFCM method was < 3%. The absolute volumes of the brain matter and cerebrospinal fluid between subjects varied quite large, ranging from 9% to 13%. The intraobserver and interobserver reproducibility for SFCM were < 4% for the soft tissues and 6% for cerebrospinal fluid. The corresponding results for the kNN segmentation method were higher compared to the SFCM method.

Quantitative MRI of the gray-white matter distribution in traumatic brain injury

Quantitative analyses were performed on magnetic resonance images (MRIs) obtained from the brains of 31 traumatic brain-injured (TBI) patients and 25 normal control subjects. The quantitative analyses involved comparisons of the shapes of proton density gray scale pixel histograms obtained from both 3-mm and 5-mm slice thickness. Image segmentation was accomplished by a multispectral fuzzy C-means and/or k-nearest-neighbor (kNN) algorithms and manual classification was used to label segmented classes into CSF, white matter, and other. Shape descriptors were derived from the pixel intensity histograms of the combined gray matter and white matter classes for each MRI slice. Statistical analyses revealed significant differences in pixel intensity distributions between TBI and control subjects. Normal control subjects tended to exhibit bimodal gray matter-white matter histograms, whereas, TBI patients tended to exhibit unimodal gray matter-white matter histograms. In the control subjects the pixels intermediate in intensity between gray and white matter were located primarily at the border between the gray and white matter, whereas TBI patients exhibited a thickening of the number of intermediate pixels at the border as well as an increase in intermediate pixels in the middle of the gray and white matter. The greater the severity of TBI, then the larger the number of intermediate intensity pixels within and between gray and white matter. Further analyses demonstrated shifts in magnetic resonance relaxation times in gray and white matter in TBI patients, which suggested that the tendency toward unimodality in TBI patients represents a pathological reduction in brain differentiation due to measurable biophysical change.