A prospective longitudinal brain morphometry study of children with sickle cell disease.

BACKGROUND AND PURPOSE: Age-related changes in brain morphology are crucial to understanding the neurobiology of sickle cell disease. We hypothesized that the growth trajectories for total GM volume, total WM volume, and regional GM volumes are altered in children with sickle cell disease compared with controls. MATERIALS AND METHODS: We analyzed T1-weighted images of the brains of 28 children with sickle cell disease (mean baseline age, 98 months; female/male ratio, 15:13) and 28 healthy age- and sex-matched controls (mean baseline age, 99 months; female/male ratio, 16:12). The total number of MR imaging examinations was 141 (2-4 for each subject with sickle cell disease, 2-3 for each control subject). Total GM volume, total WM volume, and regional GM volumes were measured by using an automated method. We used the multilevel-model-for-change approach to model growth trajectories. RESULTS: Total GM volume in subjects with sickle cell disease decreased linearly at a rate of 411 mm(3) per month. For controls, the trajectory of total GM volume was quadratic; we did not observe a significant linear decline. For subjects with sickle cell disease, we found 35 brain structures that demonstrated age-related GM volume reduction. Total WM volume in subjects with sickle cell disease increased at a rate of 452 mm(3) per month, while the trajectory of controls was quadratic. CONCLUSIONS: There was a significant age-related decrease in total GM volume in children with sickle cell disease. The GM volume reduction was spatially distributed widely across the brain, primarily in the frontal, parietal, and occipital lobes. Total WM volume in subjects with sickle cell disease increased at a lower rate than for controls.

Identification of minimal hepatic encephalopathy in patients with cirrhosis based on white matter imaging and Bayesian data mining.

BACKGROUND AND PURPOSE: White matter abnormalities have been demonstrated to play an important role in minimal hepatic encephalopathy. In this study, we aimed to evaluate whether WM diffusion tensor imaging can be used to identify minimal hepatic encephalopathy among patients with cirrhosis. MATERIALS AND METHODS: Our study included 65 patients with cirrhosis with covert hepatic encephalopathy (29 with minimal hepatic encephalopathy and 36 without hepatic encephalopathy). Participants underwent DTI, from which we generated mean diffusivity and fractional anisotropy maps. We used a Bayesian machine-learning technique, called Graphical-Model-based Multivariate Analysis, to determine WM regions that characterize group differences. To further test the clinical significance of these potential biomarkers, we performed Cox regression analysis to assess the potential of these WM regions in predicting survival. RESULTS: In mean diffusivity or fractional anisotropy maps, 2 spatially distributed WM regions (predominantly located in the bilateral frontal lobes, corpus callosum, and parietal lobes) were consistently identified as differentiating minimal hepatic encephalopathy from no hepatic encephalopathy and yielded 75.4%-81.5% and 83.1%-92.3% classification accuracy, respectively. We were able to follow 55 of 65 patients (median = 18 months), and 15 of these patients eventually died of liver-related causes. Survival analysis indicated that mean diffusivity and fractional anisotropy values in WM regions were predictive of survival, in addition to the Child-Pugh score. CONCLUSIONS: Our findings indicate that WM DTI can provide useful biomarkers differentiating minimal hepatic encephalopathy from no hepatic encephalopathy, which would be helpful for minimal hepatic encephalopathy detection and subsequent treatment.

A Bayesian diagnostic system to differentiate glioblastomas from solitary brain metastases.

This paper aimed to construct a Bayesian network-based decision support system to differentiate glioblastomas from solitary metastases, based on multimodality MR examination. We enrolled 51 patients with solitary brain tumors (26 with glioblastomas and 25 with solitary brain metastases). These patients underwent contrast-enhanced T1-weighted magnetic resonance (MR) examination, diffusion tensor imaging (DTI), dynamic susceptibility contrast (DSC) MRI, and fluid-attenuated inversion recovery (FLAIR). We generated a set of MR biomarkers, including relative cerebral blood volume in the enhancing region, and fractional anisotropy measured in the immediate peritumoral area. We then generated a Bayesian network model to represent associations among these imaging-derived predictors, and the group membership variable, (glioblastoma or solitary metastasis). This Bayesian network can be used to classify new patients' tumors based on their MR appearance. The Bayesian network model accurately differentiated glioblastomas from solitary metastases. Prediction accuracy was 0.94 (sensitivity = 0.96, specificity = 0.92) based on leave-one-out cross-validation. The area under the receiver operating characteristic curve was 0.90. A Bayesian network-based decision support system accurately differentiates glioblastomas from solitary metastases, based on MR-derived biomarkers.

Prediction of conversion from mild cognitive impairment to Alzheimer disease based on bayesian data mining with ensemble learning.

Prediction of disease progress is of great importance to Alzheimer disease (AD) researchers and clinicians. Previous attempts at constructing predictive models have been hindered by undersampling, and restriction to linear associations among variables, among other problems. To address these problems, we propose a novel Bayesian data-mining method called Bayesian Outcome Prediction with Ensemble Learning (BOPEL). BOPEL uses a Bayesian-network representation with boosting, to allow the detection of nonlinear multivariate associations, and incorporates resampling-based feature selection to prevent over-fitting caused by undersampling. We demonstrate the use of this approach in predicting conversion to AD in individuals with mild cognitive impairment (MCI), based on structural magnetic-resonance and magnetic-resonance- spectroscopy data. This study includes 26 subjects with amnestic MCI: the converter group (n = 8) met MCI criteria at baseline, but converted to AD within five years, whereas the non-converter group (n = 18) met MCI criteria at baseline and at follow-up. We found that BOPEL accurately differentiates MCI converters from non-converters, based on the baseline volumes of the left hippocampus, the banks of the right superior temporal sulcus, the right entorhinal cortex, the left lingual gyrus, and the rostral aspect of the left middle frontal gyrus. Prediction accuracy was 0.81, sensitivity was 0.63 and specificity was 0.89. We validated the generated predictive model with an independent data set constructed from the Alzheimer Disease Neuroimaging Initiative database, and again found high predictive accuracy (0.75).

Atlas-Based Classification of Hyperintense Regions from MR Diffusion-Weighted Images of the Brain: Preliminary Results.

The study of subjects with acquired brain damage in a specific location is important in exploring human brain function. Description of lesion locations within and across subjects is a crucial methodological component that usually involves the distinction of normal from damaged tissue (lesion segmentation) in relation to lesion locations in terms of a standard anatomical reference space (lesion mapping). Our study provides an atlas-based, computer-aided methodology for classification of hyperintense regions on diffusion-weighted images of the brain, representing either ischemic lesions or susceptibility artifacts. We applied a leave-one-out method of cross-validation that computed probabilistic atlases of true lesions and artifacts, based on training data. Our approach accurately classifies lesions and artifacts, but leaves a significant number of regions unclassified, due to the relatively small number of training samples. An initial segmentation step based on a larger sample of data sets is required to automate discrimination of lesions and artifacts.

Predictive models for subtypes of autism spectrum disorder based on single-nucleotide polymorphisms and magnetic resonance imaging.

PURPOSE: Autism spectrum disorder (ASD) is a neurodevelopmental disorder, of which Asperger syndrome and high-functioning autism are subtypes. Our goal is: 1) to determine whether a diagnostic model based on single-nucleotide polymorphisms (SNPs), brain regional thickness measurements, or brain regional volume measurements can distinguish Asperger syndrome from high-functioning autism; and 2) to compare the SNP, thickness, and volume-based diagnostic models. MATERIAL AND METHODS: Our study included 18 children with ASD: 13 subjects with high-functioning autism and 5 subjects with Asperger syndrome. For each child, we obtained 25 SNPs for 8 ASD-related genes; we also computed regional cortical thicknesses and volumes for 66 brain structures, based on structural magnetic resonance (MR) examination. To generate diagnostic models, we employed five machine-learning techniques: decision stump, alternating decision trees, multi-class alternating decision trees, logistic model trees, and support vector machines. RESULTS: For SNP-based classification, three decision-tree-based models performed better than the other two machine-learning models. The performance metrics for three decision-tree-based models were similar: decision stump was modestly better than the other two methods, with accuracy = 90%, sensitivity = 0.95 and specificity = 0.75. All thickness and volume-based diagnostic models performed poorly. The SNP-based diagnostic models were superior to those based on thickness and volume. For SNP-based classification, rs878960 in GABRB3 (gamma-aminobutyric acid A receptor, beta 3) was selected by all tree-based models. CONCLUSION: Our analysis demonstrated that SNP-based classification was more accurate than morphometry-based classification in ASD subtype classification. Also, we found that one SNP--rs878960 in GABRB3--distinguishes Asperger syndrome from high-functioning autism.

Automated Bayesian segmentation of microvascular white-matter lesions in the ACCORD-MIND study.

PURPOSE: Automatic brain-lesion segmentation has the potential to greatly expand the analysis of the relationships between brain function and lesion locations in large-scale epidemiologic studies, such as the ACCORD-MIND study. In this manuscript we describe the design and evaluation of a Bayesian lesion-segmentation method, with the expectation that our approach would segment white-matter brain lesions in MR images without user intervention. MATERIALS AND METHODS: Each ACCORD-MIND subject has T1-weighted, T2-weighted, spin-density-weighted, and FLAIR sequences. The training portion of our algorithm first registers training images to a standard coordinate space; then, it collects statistics that capture signal-intensity information, and residual spatial variability of normal structures and lesions. The classification portion of our algorithm then uses these statistics to segment lesions in images from new subjects, without the need for user intervention. We evaluated this algorithm using 42 subjects with primarily white-matter lesions from the ACCORD-MIND project. RESULTS: Our experiments demonstrated high classification accuracy, using an expert neuroradiologist as a standard. CONCLUSIONS: A Bayesian lesion-segmentation algorithm that collects multi-channel signal-intensity and spatial information from MR images of the brain shows potential for accurately segmenting brain lesions in images obtained from subjects not used in training.

Classification of childhood white matter disorders using proton MR spectroscopic imaging.

BACKGROUND AND PURPOSE: Childhood white matter disorders often show similar MR imaging signal-intensity changes, despite different underlying pathophysiologies. The purpose of this study was to determine if proton MR spectroscopic imaging ((1)H-MRSI) may help identify tissue pathophysiology in patients with leukoencephalopathies. MATERIALS AND METHODS: Seventy patients (mean age, 6; range, 0.66-17 years) were prospectively examined by (1)H-MRSI; a diagnosis of leukoencephalopathy due to known genetic defects leading to lack of formation, breakdown of myelin, or loss of white matter tissue attenuation (rarefaction) was made in 47 patients. The diagnosis remained undefined (UL) in 23 patients. Patients with definite diagnoses were assigned (on the basis of known pathophysiology) to 3 groups corresponding to hypomyelination, white matter rarefaction, and demyelination. Choline (Cho), creatine (Cr), and N-acetylaspartate (NAA) signals from 6 white matter regions and their intra- and intervoxel (relative to gray matter) ratios were measured. Analysis of variance was performed by diagnosis and by pathophysiology group. Stepwise linear discriminant analysis was performed to construct a model to predict pathophysiology on the basis of (1)H-MRSI, and was applied to the UL group. RESULTS: Analysis of variance by diagnosis showed 3 main metabolic patterns. Analysis of variance by pathophysiology showed significant differences for Cho/NAA (P < .001), Cho/Cr (P < .004), and NAA/Cr (P < .002). Accuracy of the linear discriminant analysis model was 75%, with Cho/Cr and NAA/Cr being the best parameters for classification. On the basis of the linear discriminant analysis model, 61% of the subjects in the UL group were classified as hypomyelinating. CONCLUSION: (1)H-MRSI provides information on tissue pathophysiology and may, therefore, be a valuable tool in the evaluation of patients with leukoencephalopathies.

Integrating data-mining support into a brain-image database using open-source components.

PURPOSE: Previously, we described our implementation of a brain-image database (braid), based on the proprietary object relational database-management system (ORDBMS) Illustra [1]. In conjunction with our collaborators, we have used this database to manage and analyze image and clinical data from what we call image-based clinical trials (IBCTs). Herein we describe the results of redesigning braid using open-source components, and integrating support for mining image and clinical data from braid's user interface. MATERIAL AND METHODS: We re-designed and re-implemented braid using open-source components, including PostgreSQL, gcc, and PHP. We integrated data-mining algorithms into braid, based on PL/R, a PostgreSQL package to support efficient communication between R and PostgreSQL. RESULTS: We present a sample clinical study to demonstrate how clinicians can perform queries for visualization, statistical analysis, and data mining, using a web-based interface. CONCLUSION: We have developed a database system with data-mining capabilities for managing, querying, analyzing and visualizing brain-MR images. We implemented this system using open-source components, with the express goal of wide dissemination throughout the neuroimaging research community.

An adaptive-focus statistical shape model for segmentation and shape modeling of 3-D brain structures.

This paper presents a deformable model for automatically segmenting brain structures from volumetric magnetic resonance (MR) images and obtaining point correspondences, using geometric and statistical information in a hierarchical scheme. Geometric information is embedded into the model via a set of affine-invariant attribute vectors, each of which characterizes the geometric structure around a point of the model from a local to a global scale. The attribute vectors, in conjunction with the deformation mechanism of the model, warranty that the model not only deforms to nearby edges, as is customary in most deformable surface models, but also that it determines point correspondences based on geometric similarity at different scales. The proposed model is adaptive in that it initially focuses on the most reliable structures of interest, and gradually shifts focus to other structures as those become closer to their respective targets and, therefore, more reliable. The proposed techniques have been used to segment boundaries of the ventricles, the caudate nucleus, and the lenticular nucleus from volumetric MR images.

Accuracy for detection of simulated lesions: comparison of fluid-attenuated inversion-recovery, proton density--weighted, and T2-weighted synthetic brain MR imaging.

OBJECTIVE: The objective of our study was to determine the effects of MR sequence (fluid-attenuated inversion-recovery [FLAIR], proton density--weighted, and T2-weighted) and of lesion location on sensitivity and specificity of lesion detection. MATERIALS AND METHODS: We generated FLAIR, proton density-weighted, and T2-weighted brain images with 3-mm lesions using published parameters for acute multiple sclerosis plaques. Each image contained from zero to five lesions that were distributed among cortical-subcortical, periventricular, and deep white matter regions; on either side; and anterior or posterior in position. We presented images of 540 lesions, distributed among 2592 image regions, to six neuroradiologists. We constructed a contingency table for image regions with lesions and another for image regions without lesions (normal). Each table included the following: the reviewer's number (1--6); the MR sequence; the side, position, and region of the lesion; and the reviewer's response (lesion present or absent [normal]). We performed chi-square and log-linear analyses. RESULTS: The FLAIR sequence yielded the highest true-positive rates (p < 0.001) and the highest true-negative rates (p < 0.001). Regions also differed in reviewers' true-positive rates (p < 0.001) and true-negative rates (p = 0.002). The true-positive rate model generated by log-linear analysis contained an additional sequence-location interaction. The true-negative rate model generated by log-linear analysis confirmed these associations, but no higher order interactions were added. CONCLUSION: We developed software with which we can generate brain images of a wide range of pulse sequences and that allows us to specify the location, size, shape, and intrinsic characteristics of simulated lesions. We found that the use of FLAIR sequences increases detection accuracy for cortical-subcortical and periventricular lesions over that associated with proton density- and T2-weighted sequences.

A simulator for evaluating methods for the detection of lesion-deficit associations.

Although much has been learned about the functional organization of the human brain through lesion-deficit analysis, the variety of statistical and image-processing methods developed for this purpose precludes a closed-form analysis of the statistical power of these systems. Therefore, we developed a lesion-deficit simulator (LDS), which generates artificial subjects, each of which consists of a set of functional deficits, and a brain image with lesions; the deficits and lesions conform to predefined distributions. We used probability distributions to model the number, sizes, and spatial distribution of lesions, to model the structure-function associations, and to model registration error. We used the LDS to evaluate, as examples, the effects of the complexities and strengths of lesion-deficit associations, and of registration error, on the power of lesion-deficit analysis. We measured the numbers of recovered associations from these simulated data, as a function of the number of subjects analyzed, the strengths and number of associations in the statistical model, the number of structures associated with a particular function, and the prior probabilities of structures being abnormal. The number of subjects required to recover the simulated lesion-deficit associations was found to have an inverse relationship to the strength of associations, and to the smallest probability in the structure-function model. The number of structures associated with a particular function (i.e., the complexity of associations) had a much greater effect on the performance of the analysis method than did the total number of associations. We also found that registration error of 5 mm or less reduces the number of associations discovered by approximately 13% compared to perfect registration. The LDS provides a flexible framework for evaluating many aspects of lesion-deficit analysis.

Quantitative 1H MR spectroscopic imaging in early Rett syndrome.

OBJECTIVE: To determine cerebral regional concentrations of N-acetyl aspartate (NAA), total choline (Cho), and total creatine (Cr) in Rett syndrome (RS) using 1H magnetic resonance spectroscopic imaging (MRSI). BACKGROUND: The biochemical defect underlying RS is unknown. Because in vivo MRSI can detect important cerebral metabolites, MRSI has a potential to reveal impairment of regional cerebral metabolism in RS noninvasively. METHODS: High-resolution, multislice 1H MRSI was carried out in 17 girls with RS. The control group consisted of nine healthy children. RESULTS: In patients with RS, average Cho concentration was 12% higher (p < 0.005) and average NAA concentration 11% lower (p < 0.0001) compared with the control group. Regional metabolic differences included significantly lower NAA concentration in the frontal gray and white matter, insula, and hippocampus in RS; no difference in regional Cho and Cr concentrations were found. A 20 to 38% higher Cho:NAA ratio in frontal and parietal gray and white matter, insular gray matter, and hippocampus (p < 0.05) and a 14 to 47% lower NAA:Cr ratio in frontal cortical gray matter, parietal and temporal white matter, insula, and putamen (p < 0.05) were found in subjects with RS compared with controls. Patients with seizures had higher average concentrations of Cho, Cr, and NAA compared with those without seizures (8-19%, p < 0.05). CONCLUSION: Metabolic impairment in RS involves both gray and white matter and particularly involves frontal and parietal lobes and the insular cortex. Loss of NAA most likely reflects reduced neuronal and dendritic tree size; increased Cho concentration may result from gliosis.

An architecture for a brain-image database.

The widespread availability of methods for noninvasive assessment of brain structure has enabled researchers to investigate neuroimaging correlates of normal aging, cerebrovascular disease, and other processes; we designate such studies as image-based clinical trials (IBCTs). We propose an architecture for a brain-image database, which integrates image processing and statistical operators, and thus supports the implementation and analysis of IBCTs. The implementation of this architecture is described and results from the analysis of image and clinical data from two IBCTs are presented. We expect that systems such as this will play a central role in the management and analysis of complex research data sets.

Is the spatial distribution of brain lesions associated with closed-head injury predictive of subsequent development of attention-deficit/hyperactivity disorder? Analysis with brain-image database.

PURPOSE: To determine whether there is an association between the spatial distribution of lesions detected at magnetic resonance (MR) imaging of the brain in children after closed-head injury and the development of secondary attention-deficit/hyperactivity disorder (ADHD). MATERIALS AND METHODS: Data obtained from 76 children without prior history of ADHD were analyzed. MR images were obtained 3 months after closed-head injury. After manual delineation of lesions, images were registered to the Talairach coordinate system. For each subject, registered images and secondary ADHD status were integrated into a brain-image database, which contains depiction (visualization) and statistical analysis software. Using this database, we assessed visually the spatial distributions of lesions and performed statistical analysis of image and clinical variables. RESULTS: Of the 76 children, 15 developed secondary ADHD. Depiction of the data suggested that children who developed secondary ADHD had more lesions in the right putamen than children who did not develop secondary ADHD; this impression was confirmed statistically. After Bonferroni correction, we could not demonstrate significant differences between secondary ADHD status and lesion burdens for the right caudate nucleus or the right globus pallidus. CONCLUSION: Closed-head injury-induced lesions in the right putamen in children are associated with subsequent development of secondary ADHD. Depiction software is useful in guiding statistical analysis of image data.

A brain image database for structure/function analysis.

BACKGROUND AND PURPOSE: Lesion-deficit-based structure-function analysis has traditionally been empirical and nonquantitative. Our purpose was to establish a new brain image database (BRAID) that allows the statistical correlation of brain functional measures with anatomic lesions revealed by clinical brain images. METHODS: Data on 303 participants in the MR Feasibility Study of the Cardiovascular Health Study were tested for lesion/deficit correlations. Functional data were derived from a limited neurologic examination performed at the time of the MR examination. Image data included 3D lesion descriptions derived from the MR examinations by hand segmentation. MR images were normalized in-plane using local, linear Talairach normalization. A database was implemented to support spatial data structures and associated geometric and statistical operations. The database stored the segmented lesions, patient functional scores, and several anatomic atlases. Lesion-deficit association was sought by contingency testing (chi2-test) for every possible combination of each neurologic variable and each labeled atlas structure. Significant associations that confirmed accepted lesion-deficit relationships were sought. RESULTS: Two-hundred thirty-five infarctlike lesions in 117 subjects were viewed collectively after mapping into Talairach cartesian coordinates. Anatomic structures most strongly correlated with neurologic deficits tended to be situated in anatomically appropriate areas. For example, infarctlike lesions associated with visual field defects were correlated with structures in contralateral occipital structures, including the optic radiations and occipital gyri. CONCLUSION: Known lesion-deficit correlations can be established by a database using a standard coordinate system for representing spatial data and incorporating functional and structural data together with appropriate query mechanisms. Improvements and further applications of this methodology may provide a powerful technique for uncovering new structure-function relationships.

Algorithms for Bayesian belief-network precomputation.

Bayesian belief networks provide an intuitive and concise means of representing probabilistic relationships among the variables in expert systems. A major drawback to this methodology is its computational complexity. We present an introduction to belief networks, and describe methods for precomputing, or caching, part of a belief network based on metrics of probability and expected utility. These algorithms are examples of a general method for decreasing expected running time for probabilistic inference. We first present the necessary background, and then present algorithms for producing caches based on metrics of expected probability and expected utility. We show how these algorithms can be applied to a moderately complex belief network, and present directions for future research.