Simon Duchesne

Whole-brain voxel-based statistical analysis of gray matter and white matter in temporal lobe epilepsy

Volumetric MRI studies based on manual labeling of selected anatomical structures have provided in vivo evidence that brain abnormalities associated with temporal lobe epilepsy (TLE) extend beyond the hippocampus. Voxel-based morphometry (VBM) is a fully automated image analysis technique allowing identification of regional differences in gray matter (GM) and white matter (WM) between groups of subjects without a prior region of interest. The purpose of this study was to determine whole-brain GM and WM changes in TLE and to investigate the relationship between these abnormalities and clinical parameters. We studied 85 patients with pharmacologically intractable TLE and unilateral hippocampal atrophy and 47 age- and sex-matched healthy control subjects. The seizure focus was right sided in 40 patients and left sided in 45. Students t test statistical maps of differences between patients and controls GM and WM concentrations were obtained using a general linear model. A further regression against duration of epilepsy, age of onset, presence of febrile convulsions, and secondary generalized seizures was performed with the TLE population. Voxel-based morphometry revealed that GM pathology in TLE extends beyond the hippocampus involving other limbic areas such as the cingulum and the thalamus, as well as extralimbic areas, particularly the frontal lobe. White matter reduction was found only ipsilateral to the seizure focus, including the temporopolar, entorhinal, and perirhinal areas. This pattern of structural changes is suggestive of disconnection involving preferentially frontolimbic pathways in patients with pharmacologically intractable TLE.

Appearance-based segmentation of medial temporal lobe structures.

A new paradigm for the characterization of structure appearance is proposed, based on a combination of gray-level MRI intensity data and a shape descriptor derived from a priori principal components analysis of 3D deformation vector fields. Generated without external intervention, it extends into 3D more classic, 2D manual landmark-based shape models. Application of this novel concept led to a method for the segmentation of medial temporal lobe structures from brain magnetic resonance images. The strategy employed for segmentation aims at synthesizing, using the appearance model, a deformation field that maps a new volume onto a reference target. Any information defined on the reference can then be propagated back on the new volume instance, thereby achieving segmentation. The proposed method was tested on a data set of 80 normal subjects and compared against manual segmentation as well as automated segmentation results from ANIMAL, a nonlinear registration and segmentation technique. Experimental results demonstrated the robustness and flexibility of the new method. Segmentation accuracy, measured by overlap statistics, is marginally lower (< 2%) than ANIMAL, while processing time is six times faster. Finally, the applicability of this concept toward shape deformation analysis is presented.

MR-based neurological disease classification methodology: application to lateralization of seizure focus in temporal lobe epilepsy.

Classification approaches for neurological diseases tend to concentrate on specific structures such as the hippocampus (HC). The hypothesis for the novel methodology presented in this work is that pathologies will impact large tissue areas with detectable variations of T1-weighted MR signal intensity and registration metrics. The technique is applied to lateralization of seizure focus in 127 patients with intractable temporal lobe epilepsy (TLE), in which the site of seizure onset was determined by comprehensive evaluation (69 with left MTL seizure focus (SF) (group L) and 58 with right SF (group R)). The method analyses large, non-specific Volumes of Interest (VOI) centered on the left and right medial temporal lobes (MTL) (55 x 82 x 80 voxels) in pre-processed scans aligned in stereotaxic space. Extracted VOIs are linearly and nonlinearly registered to a reference target image. Principal Components Analyses of (i) the normalized intensity and (ii) the trace, a measure of local volume change, are used to generate a multidimensional reference space from a set of 152 neurologically healthy subjects. VOIs from TLE patients, processed in a similar fashion, are projected in this space, and leave-one-out, forward stepwise linear discriminant analysis of the eigencoordinate distributions is used for classification. Following manual MRI volumetric analysis, 80 patients had HC atrophy (group HA) ipsilateral to the SF (42 with left SF or LHA, and 38 with right or RHA), and the remaining 47 had normal HC volumes (group HNV) (27 with left SF or LNV, and 20 with right SF or RNV). The automated method was 100% accurate at separating HA vs. HNV, LHA vs. RHA, and LNV vs RNV. It was also 96% accurate at separating L vs. R. Our results indicate that MR data projected in multidimensional feature domains can lateralize SF in epilepsy patients with a high accuracy, irrespective of HC volumes. This single-scan, practical and objective method holds promise for the pre-surgical evaluation of TLE patients.

Predicting clinical variable from MRI features: application to MMSE in MCI

The ability to predict a clinical variable from automated analysis of single, cross-sectional T1-weighted (T1w) MR scans stands to improve the management of patients with neurological diseases. We present a methodology for predicting yearly Mini-Mental Score Examination (MMSE) changes in Mild Cognitive Impairment (MCI) patients. We begin by generating a non-pathological, multidimensional reference space from a group of 152 healthy volunteers by Principal Component Analyses of (i) T1w MR intensity of linearly registered Volumes of Interest (VOI); and (ii) trace of the deformation fields of nonlinearly registered VOIs. We use multiple regression to build linear models from eigenvectors where the projection eigencoordinates of patient data in the reference space are highly correlated with the clinical variable of interest. In our cohort of 47 MCI patients, composed of 16 decliners, 26 stable and 5 improvers (based on MMSE at 1 yr follow-up), there was a significant difference (P = 0.0003) for baseline MMSE scores between decliners and improvers, but no other differences based on age or sex. First, we classified our three groups using leave-one-out, forward stepwise linear discriminant analyses of the projection eigencoordinates with 100% accuracy. Next, we compared various linear models by computing F-statistics on the residuals of predicted vs actual values. The best model was based on 10 eigenvectors + baseline MMSE, with predicted yearly changes highly correlated (r = 0.6955) with actual data. Prospective study of an independent cohort of patients is the next logical step towards establishing this promising technique for clinical use.

Assessment of adolescent body perception: Development and characterization of a novel tool for morphing images of adolescent bodies

We developed a computer-based method of distorting adolescent body images, which incorporates the covariation between body parts found during growth and sexual maturation. An Adolescent Body-Shape Database (AdoBSD) and Adolescent Body Morphing Tool (AdoBMT) are described; the AdoBSD comprises real (n = 320) and morphed (n ≈ 41,000) images (front and side view) of 160 adolescents (9–17 years). We used a point distribution model, based upon principal components analysis, to characterize the covariation between predefined body tag-points manually positioned on the body images and to morph the body images in a realistic manner. Eight principal components (PCs) were found to characterize 96.3% of the covariation between body tag-points. Application of the PCs to the body images resulted in the manipulation of body parts including shoulder width, waist, hip, belly, thigh and calf sizes. The AdoBMT and AdoBSD may be used to investigate changes in body perception during adolescence, and the role of body perception in adolescent obesity and eating disorders. The AdoBSD is available to the research community (www.brainbody.nottingham.ac.uk).

On the classification of temporal lobe epilepsy using MR image appearance

Classification of neurological diseases based on image characteristics often requires extensive modeling and user intervention. While other techniques concentrate on specific structures, the novelty of the method presented here resides in its analysis of the grey-level appearance of large, non-specific Volumes of Interest (VOI) from T1 MRI data. No manual intervention is required other than the selection of the VOI. This work presents the methodological framework and preliminary results towards our aim of classifying normal subjects and patients with Temporal Lobe Epilepsy (TLE) within the Medial Temporal Lobe. For this purpose, principal component analysis is performed on a set of normal subjects for the creation of a multi-dimensional space representative of a normal population. New data for normal and TLE subjects are projected in this space, under the assumption that the distributions of the projections are not identical and can be used for classification. It is shown that linear discriminant analysis of the eigencoordinates of the projected data can be used to classify normals vs TLE with a 70% accuracy based on only 10 eigenvectors. This results can go up to 100% if all eigenvectors defining the grey-level space are used.

Appearance-based modelling and segmentation of the hippocampus from MR images

Current segmentation techniques of the hippocampus from MR images generally require manual intervention or extensive computation time. Not all methods incorporate statistical information on the structure or volume of interest. This work is novel in that it presents a fully 3D, non-supervised appearance-based method for segmentation, hippocampus, based on a priori analysis of deformation fields. Early segmentation results demonstrate that this method is as accurate as ANIMAL, a non-linear registration and segmentation technique, while being faster. Refinements in the training strategy of the model should further improve accuracy with no additional on-line computational expense. A key feature of this approach is its ability to segment other structures of interest simply by retraining the model off-line on a new data set. The applicability of the proposed model towards shape deformation analysis is discussed.

Successful AD and MCI differentiation from normal aging via automated analysis of MR image features

slice thickness 2mm, 860 m pixel-size, for a total imaging time of 6 minutes for four images. Each pixel’s signal intensity was fitted as a function of TSL by a linear least-squares algorithm to generate T1 maps. A region of interest was manually selected in the parenchyma in the right medial temporal lobe and average T1 values recorded. A student’s t-test was performed to determine any significance in the difference in T1 values between groups. Conclusions: Average T1 for the AD group was 95.3 1.4ms (mean std. error) and for controls was 87.5 1.7ms and the difference was statistically significant (p 0.005). Typical T1 MR images (grayscale images) and corresponding T1 maps (color images) are shownin the figure. This is the first demonstration of imaging AD pathology in humans using T1 imaging. Our results indicate that clinically-confirmed AD results in the prolongation of T1 relaxation time in brain tissue.

Temporal Lobe Epilepsy Lateralization Based on MR Image Intensity and Registration Features

In the context of MR imaging, explicit segmentation followed by stereologic volumetry of the hippocampus (HC) has been the standard approach toward temporal lobe epilepsy (TLE) lateralization of the seizure focus. The novelty of the method presented here resides in its analysis of characteristics of large, non-specific Volumes of Interest from T1 MRI data aiming to lateralize the seizure focus in patients with TLE without segmentation. For this purpose, Principal Components Analysis (PCA) of two image features are united to create a multi-dimensional space representative of a training set population composed of 150 normal subjects. The feature instances consist of grey-level intensity and an approximation of the Jacobian matrix of non-linear registration-derived dense deformation fields. New data for TLE subjects are projected in this space, under the assumption that the distributions of the projections of normal and patients are not identical and can be used for lateralization. Results are presented following PCA modeling of the left medial temporal lobe only for all subjects. It is shown that linear discriminant analysis of the eigencoordinates can be used to lateralize the seizure focus in TLE patients with a 75% accuracy. It is expected that adding a right temporal lobe model will improve lateralization results beyond those of HC volumetry.

Analysis of 3D Deformation Fields for Appearance-Based Segmentation

Segmentation methods for brain MR images typically employ manual and/or automatic knowledge-based models specific to the structure of interest (SOI). The technique presented here overcomes some of the limitations of current methods. It requires no manual intervention, is fast, fully 3D, and generic yet constrained by some form of prior structure information. The novelty of this work resides in its a priori Principal Components Analysis (PCA) of non-linear registration data of a volume of interest (VOI), represented by dense 3D deformation fields from ANIMAL [1]. The results are used in an Appearance Model, inspired by Cootes [2], able to segment any SOIs contained within the VOI, in the atlas-independent framework described by Collins [1]. This article presents the theoretical basis for and initial work towards hippocampus segmentation on subject images from the MNI International Consortium for Brain Mapping (ICBM) database.