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

Multispectral magnetic resonance images segmentation using fuzzy Hopfield neural network

This paper demonstrates a fuzzy Hopfield neural network for segmenting multispectral MR brain images. The proposed approach is a new unsupervised 2-D Hopfield neural network based upon the fuzzy clustering technique. Its implementation consists of the combination of 2-D Hopfield neural network and fuzzy c-means clustering algorithm in order to make parallel implementation for segmenting multispectral MR brain images feasible. For generating feasible results, a fuzzy c-means clustering strategy is included in the Hopfield neural network to eliminate the need for finding weighting factors in the energy function which is formulated and based on a basic concept commonly used in pattern classification, called the 'within-class scatter matrix' principle. The suggested fuzzy c-means clustering strategy has also been proven to be convergent and to allow the network to learn more effectively than the conventional Hopfield neural network. The experimental results show that a near optimal solution can be obtained using the fuzzy Hopfield neural network based on the within-class scatter matrix.

Post-acquisition correction of MR inhomogeneities

Signal inhomogeneities in volumetric head MR scans are a major obstacle to segmentation and neuromorphometry. The fuzzy c-means (FCM) statistical clustering algorithm was extended to estimate and retrospectively correct a multiplicative inhomogeneity field in T1-weighted head MR scans. The method was tested on a mathematically simulated object and on seven whole head 3D MR scans. Once initial parameters governing operation of the algorithm were chosen for this class of images, results were obtained without intervention for individual MR studies. Post-acquisition inhomogeneity correction by extended FCM clustering improved overall image uniformity and separability of gray and white matter intensities.

The application of competitive Hopfield neural network to medical image segmentation

In this paper, a parallel and unsupervised approach using the competitive Hopfield neural network (CHNN) is proposed for medical image segmentation. It is a kind of Hopfield network which incorporates the winner-takes-all (WTA) learning mechanism. The image segmentation is conceptually formulated as a problem of pixel clustering based upon the global information of the gray level distribution. Thus, the energy function for minimization is defined as the mean of the squared distance measures of the gray levels within each class. The proposed network avoids the onerous procedure of determining values for the weighting factors in the energy function. In addition, its training scheme enables the network to learn rapidly and effectively. For an image of n gray levels and c interesting objects, the proposed CHNN would consist of n by c neurons and be independent of the image size. In both simulation studies and practical medical image segmentation, the CHNN method shows promising results in comparison with two well-known methods: the hard and the fuzzy c-means (FCM) methods.

Knowledge-based interpretation of MR brain images

The authors have developed a method for fully automated segmentation and labeling of 17 neuroanatomic structures such as thalamus, caudate nucleus, ventricular system, etc. in magnetic resonance (MR) brain images. The authors' method is based on a hypothesize-and-verify principle and uses a genetic algorithm (GA) optimization technique to generate and evaluate image interpretation hypotheses in a feedback loop. The authors' method was trained in 20 individual T1-weighted MR images. Observer-defined contours of neuroanatomic structures were used as a priori knowledge. The method's performance was validated in eight MR images by comparison to observer-defined independent standards. The GA-based image interpretation method correctly interpreted neuroanatomic structures in all images from the test set. Computer-identified and observer-defined neuroanatomic structure areas correlated very well (r=0.99, y=0,95x-2.1). Border positioning errors were small, with a root mean square (rms) border positioning error of 1.5+/-0.6 pixels. The authors' GA-based image interpretation method represents a novel approach to image interpretation and has been shown to produce accurate labeling of neuroanatomic structures in a set of MR brain images.

Labeling of MR brain images using Boolean neural network

Presents a knowledge-based approach for labeling two-dimensional (2-D) magnetic resonance (MR) brain images using the Boolean neural network (BNN), which has binary inputs and outputs, integer weights, fast learning and classification, and guaranteed convergence. The approach consists of two components: a BNN clustering algorithm and a constraint satisfying Boolean neural network (CSBNN) labeling procedure. The BNN clustering algorithm is developed to initially segment an image into a number of regions. Then the segmented regions are labeled with the CSBNN, which is a modified version of BNN. The CSBNN uses a knowledge base that contains information on image-feature space and tissue models as constraints. The method is tested using sets of MR brain images. The regions of the different brain tissues are satisfactorily segmented and labeled. A comparison with the Hopfield neural network and the traditional simulated annealing method for image labeling is provided. The comparison results show that the CSBNN approach offers a fast, feasible, and reliable alternative to the existing techniques for medical image labeling.

A technique for single-channel MR brain tissue segmentation: application to a pediatric sample

A segmentation method is presented for gray matter, white matter, and cerebrospinal fluid (CSF) in thin-sliced single-channel brain magnetic resonance (MR) scans. The method is based on probabilistic modeling of intensity distributions and on a region growing technique. Interrater and intrarater reliabilities for the method were high, and comparison with phantom studies and hand-traced results from an experienced rater indicated good validity. The method was designed to account for spatially dependent image intensity inhomogeneities. Segmentation of MR brain scans of 105 (56 male and 49 female) healthy children and adolescents showed that although the total brain volume was stable over age 4-18, white matter increased and gray matter decreased significantly. There were no sex differences in total gray and white matter growth after correction for total brain volume. White matter volume increased the most in superior and posterior regions and laterality effects were seen in hemisphere tissue volumes. These findings are consistent with other reports, and further validate the segmentation technique.

Classification of brain compartments and head injury lesions by neural networks applied to MRI

An automatic, neural network-based approach was applied to segment normal brain compartments and lesions on MR images. Two supervised networks, backpropagation (BPN) and counterpropagation, and two unsupervised networks, Kohonen learning vector quantizer and analog adaptive resonance theory, were trained on registered T2-weighted and proton density images. The classes of interest were background, gray matter, white matter, cerebrospinal fluid, macrocystic encephalomalacia, gliosis, and "unknown." A comprehensive feature vector was chosen to discriminate these classes. The BPN combined with feature conditioning, multiple discriminant analysis followed by Hotelling transform, produced the most accurate and consistent classification results. Classification of normal brain compartments were generally in agreement with expert interpretation of the images. Macrocystic encephalomalacia and gliosis were recognized and, except around the periphery, classified in agreement with the clinician's report used to train the neural network.

Unsupervised measurement of brain tumor volume on MR images

We examined unsupervised methods of segmentation of MR images of the brain for measuring tumor volume in response to treatment. Two clustering methods were used: fuzzy c-means and a nonfuzzy clustering algorithm. Results were compared with volume segmentations by two supervised methods, k-nearest neighbors and region growing, and all results were compared with manual labelings. Results of individual segmentations are presented as well as comparisons on the application of the different methods with 10 data sets of patients with brain tumors. Unsupervised segmentation is preferred for measuring tumor volumes in response to treatment, as it eliminates operator dependency and may be adequate for delineation of the target volume in radiation therapy. Some obstacles need to be overcome, in particular regarding the detection of anatomically relevant tissue classes. This study shows that these improvements are possible.

Automatic magnetic resonance tissue characterization for three-dimensional magnetic resonance imaging of the brain

Computer-assisted diagnostic systems enhance the information available from magnetic resonance imaging. Segmentations are the basis on which three-dimensional volume renderings are made. The application of a raw data-based, operator-independent (automatic), magnetic resonance segmentation technique for tissue differentiation is demonstrated. Segmentation images of vasogenic edema with gross and histopathological correlation are presented for demonstration of the technique. A pixel was classified into a tissue class based on a feature vector using unsupervised fuzzy clustering techniques as the pattern recognition method. Correlation of fuzzy segmentations and gross and histopathology were successfully performed. Based on the results of neuropathological correlation, the application of fuzzy magnetic resonance image segmentation to a patient with a brain tumor and extensive edema represents a viable technique for automatically displaying clinically important tissue differentiation. With this pattern recognition technique, it is possible to generate automatic segmentation images that display diagnostically relevant neuroanatomical and neuropathological tissue contrast information from raw magnetic resonance data for use in three-dimensional volume reconstructions.

MRI segmentation: methods and applications

The current literature on MRI segmentation methods is reviewed. Particular emphasis is placed on the relative merits of single image versus multispectral segmentation, and supervised versus unsupervised segmentation methods. Image pre-processing and registration are discussed, as well as methods of validation. The application of MRI segmentation for tumor volume measurements during the course of therapy is presented here as an example, illustrating problems associated with inter- and intra-observer variations inherent to supervised methods.

Comparison of supervised MRI segmentation methods for tumor volume determination during therapy

Two different multispectral pattern recognition methods are used to segment magnetic resonance images (MRI) of the brain for quantitative estimation of tumor volume and volume changes with therapy. A supervised k-nearest neighbor (kNN) rule and a semi-supervised fuzzy c-means (SFCM) method are used to segment MRI slice data. Tumor volumes as determined by the kNN and SFCM segmentation methods are compared with two reference methods, based on image grey scale, as a basis for an estimation of ground truth, namely: (a) a commonly used seed growing method that is applied to the contrast enhanced T1-weighted image, and (b) a manual segmentation method using a custom-designed graphical user interface applied to the same raw image (T1-weighted) dataset. Emphasis is placed on measurement of intra and inter observer reproducibility using the proposed methods. Intra- and interobserver variation for the kNN method was 9% and 5%, respectively. The results for the SFCM method was a little better at 6% and 4%, respectively. For the seed growing method, the intra-observer variation was 6% and the interobserver variation was 17%, significantly larger when compared with the multispectral methods. The absolute tumor volume determined by the multispectral segmentation methods was consistently smaller than that observed for the reference methods. The results of this study are found to be very patient case-dependent. The results for SFCM suggest that it should be useful for relative measurements of tumor volume during therapy, but further studies are required. This work demonstrates the need for minimally supervised or unsupervised methods for tumor volume measurements.

Application of fuzzy c-means segmentation technique for tissue differentiation in MR images of a hemorrhagic glioblastoma multiforme

The application of a raw data-based, operator-independent MR segmentation technique to differentiate boundaries of tumor from edema or hemorrhage is demonstrated. A case of a glioblastoma multiforme with gross and histopathologic correlation is presented. The MR image data set was segmented into tissue classes based on three different MR weighted image parameters (T1-, proton density-, and T2-weighted) using unsupervised fuzzy c-means (FCM) clustering algorithm technique for pattern recognition. A radiological examination of the MR images and correlation with fuzzy clustering segmentations was performed. Results were confirmed by gross and histopathology which, to the best of our knowledge, reports the first application of this demanding approach. Based on the results of neuropathologic correlation, the application of FCM MR image segmentation to several MR images of a glioblastoma multiforme represents a viable technique for displaying diagnostically relevant tissue contrast information used in 3D volume reconstruction. With this technique, it is possible to generate segmentation images that display clinically important neuroanatomic and neuropathologic tissue contrast information from raw MR image data.

Head surface digitization and registration: a method for mapping positions on the head onto magnetic resonance images

We have developed a method for mapping positions on the head, such as anatomical landmarks, electrode locations, and stimulation sites, onto magnetic resonance (MR) images of the head. This method is based on the registration of two representations of the head surface: a series of contours obtained from MR images and a set of points measured from the head. The three-dimensional coordinates of each head point were acquired with the use of a magnetic digitizer, whose source was removed from the equipment and mounted on top of the subject's head. This arrangement seemed less uncomfortable for the subject than head immobilization and allowed the acquisition of many points without compromising the precision of the measurements. The digitized head surface was registered to MR image head contours using a surface registration algorithm. The registration provided the rotation and translation parameters needed for mapping head positions onto MR images. The precision of this mapping method has been estimated to be in the range of 3 to 8 mm. This method has been used to map dipole sources in electroencephalography and magneto-encephalography and to impose maps of scalp sites used in transcranial magnetic stimulation onto MR and PET images of the brain.

Estimation of CSF, white and gray matter volumes in hydrocephalic children using fuzzy clustering of MR images

An algorithm and set of procedures for measuring volumes of cerebrospinal fluid (CSF), white matter, and gray matter from transaxial magnetic resonance images (MRI) of the brain are described. The algorithm is a variant of the fuzzy c-means clustering method for texture identification. This technique is used mainly to solve the problem of volume averaging of tissue compartments, but also has other advantages. It is fast, accurate, and relatively operator independent. Furthermore, it does not depend on statistical assumptions such as data normality, nor does it require any a priori heuristics. The procedure was tested successfully on imaged phantoms of known volume composition and compared with results achieved using a standard morphometric measurement approach. The procedure was also applied to brain MRIs of three clinically normal children and three age-matched children with hydrocephalus using both proton density and T2-weighted images. The algorithm was able to detect the expected increased amounts of CSF and decreased amounts of white matter characteristic of the hydrocephalic brain.

Parameter estimation and tissue segmentation from multispectral MR images

A statistical method is developed to classify tissue types and to segment the corresponding tissue regions from relaxation time T(1 ), T(2), and proton density P(D) weighted magnetic resonance images. The method assumes that the distribution of image intensities associated with each tissue type can be expressed as a multivariate likelihood function of three weighted signal intensity values (T(1), T(2), P(D)) at each location within that tissue regions. The method further assumes that the underlying tissue regions are piecewise contiguous and can be characterized by a Markov random field prior. In classifying the tissue types, the method models the likelihood of realizing the images as a finite multivariate-mixture function. The class parameters associated with the tissue types (i.e. the weighted intensity means, variances and correlation coefficients of the multivariate function, as well as the number of voxels within regions of the tissue types of are estimated by maximum likelihood. The estimation fits the class parameters to the image data via the expectation-maximization algorithm. The number of classes associated with the tissue types is determined by the information criterion of minimum description length. The method segments the tissue regions, given the estimated class parameters, by maximum a posteriori probability. The prior is constructed by the tissue-region membership of the first- and second-order neighborhood. The method is tested by a few sets of T(1), T(2), and P(D) weighted images of the brain acquired with a 1.5 Tesla whole body scanner. The number of classes and the associated class parameters are automatically estimated. The regions of different brain tissues are satisfactorily segmented.

A visual fuzzy cluster system for patient analysis

A visual fuzzy cluster (VFC) system is developed to assist physicians in interactively detecting and refining cluster partitions for patients with acute upper respiratory infections. The VFC system assists physicians in discovering relationships among patients by applying a fuzzy cluster algorithm to analyse a case base of patient findings. The algorithm discovers similarities among patients, while at the same time identifying atypical patients. The system visually presents the fuzzy cluster solutions on a three-dimensional animated display. Physicians then interactively manipulate icons representing patients to explore and refine the fuzzy cluster solution. Initial experiences with the VFC prototype are encouraging and support the claims that the system improves physician understanding and allows physicians to take advantage of visual recognition and manipulation skills to define and label patient groupings. The resulting labels and cluster centres or prototypes offer insight into the set of patient features that best discriminate between groups.

Segmentation of medical images through competitive learning

In image analysis applications, segmentation of gray-level images into meaningful regions is an important low-level processing step. Various approaches to segmentation investigated in the literature, in general, use either local information of gray-level values of pixels (region growing based methods, for example) or the global information (histogram thresholding based methods, for example). Application of these approaches for segmenting medical images often does not provide satisfactory results. Medical images are usually characterized by low local contrast and noisy or faded features causing unacceptable performance of local information based segmentation methods. In addition, because of a large amount of structural information found in medical images, global information based segmentation methods yield inadequate results in region extraction. We present a novel approach to image segmentation that combines local contrast as well as global gray-level distribution information. The presented method adaptively learns useful features and regions through the use of a normalized contrast function as a measure of local information and a competitive learning based method to update region segmentation incorporating global information about the gray-level distribution of the image. In this paper, we present the framework of such a self organizing feature map, and show the results on simulated as well as real medical images.

Knowledge-based classification and tissue labeling of MR images of human brain

Presents a knowledge-based approach to automatic classification and tissue labeling of 2D magnetic resonance (MR) images of the human brain. The system consists of 2 components: an unsupervised clustering algorithm and an expert system. MR brain data is initially segmented by the unsupervised algorithm, then the expert system locates a landmark tissue or cluster and analyzes it by matching it with a model or searching in it for an expected feature. The landmark tissue location and its analysis are repeated until a tumor is found or all tissues are labeled. The knowledge base contains information on cluster distribution in feature space and tissue models. Since tissue shapes are irregular, their models and matching are specially designed: 1) qualitative tissue models are defined for brain tissues such as white matter; 2) default reasoning is used to match a model with an MR image; that is, if there is no mismatch between a model and an image, they are taken as matched. The system has been tested with 53 slices of MR images acquired at different times by 2 different scanners. It accurately identifies abnormal slices and provides a partial labeling of the tissues. It provides an accurate complete labeling of all normal tissues in the absence of large amounts of data nonuniformity, as verified by radiologists. Thus the system can be used to provide automatic screening of slices for abnormality. It also provides a first step toward the complete description of abnormal images for use in automatic tumor volume determination.

Correction of intensity variations in MR images for computer-aided tissue classification

A number of supervised and unsupervised pattern recognition techniques have been proposed in recent years for the segmentation and the quantitative analysis of MR images. However, the efficacy of these techniques is affected by acquisition artifacts such as inter-slice, intra-slice, and inter-patient intensity variations. Here a new approach to the correction of intra-slice intensity variations is presented. Results demonstrate that the correction process enhances the performance of backpropagation neural network classifiers designed for the segmentation of the images. Two slightly different versions of the method are presented. The first version fits an intensity correction surface directly to reference points selected by the user in the images. The second version fits the surface to reference points obtained by an intermediate classification operation. Qualitative and quantitative evaluation of both methods reveals that the first one leads to a better correction of the images than the second but that it is more sensitive to operator errors.

MRI: stability of three supervised segmentation techniques

Supervised segmentation methods from three families of pattern recognition techniques were used to segment multispectral MRI data. Studied were the maximum likelihood method (MLM), k-nearest neighbors (k-NN), and a back-propagation artificial neural net (ANN). Performance was measured in terms of execution speed, and stability for the selection of training data, namely, region of interest (ROI) selection, and interslice and interpatient classifications. MLM proved to have the smallest execution times, but demonstrated the least stability. k-NN showed the best stability for training data selection. To evaluate the segmentation techniques, multispectral images were used of normal volunteers and patients with gliomas, the latter with and without MR contrast material. All measures applied indicated that k-NN provides the best results.

Neural-network-based segmentation of multi-modal medical images: a comparative and prospective study

This work presents an investigation of the potential of artificial neural networks for classification of registered magnetic resonance and X-ray computer tomography images of the human brain. First, topological and learning parameters are established experimentally. Second, the learning and generalization properties of the neural networks are compared to those of a classical maximum likelihood classifier and the superiority of the neural network approach is demonstrated when small training sets are utilized. Third, the generalization properties of the neural networks are utilized to develop an adaptive learning scheme able to overcome interslice intensity variations typical of MR images. This approach permits the segmentation of image volumes based on training sets selected on a single slice. Finally, the segmentation results obtained both with the artificial neural network and the maximum likelihood classifiers are compared to contours drawn manually.