• voxel-based morphometry
• support vector machines
• neuroimaging
• autism spectrum disorders

This thesis aims to examine two types of elaboration techniques of brain magnetic resonance imaging (MRI) data: the voxel-based morphometry (VBM) and the support vector machine (SVM) approaches. While the VBM is a standard and well-established mass-univariate
method, the SVM multivariate analysis has been rarely implemented to investigate brain MRI
data. An improvement of our knowledge on the pattern classication approach is necessary
to be achieved, both to assess its exploratory capability and to point out advantages and disadvantages with respect to the more largely used VBM approach. Despite these methods are
potentially suitable to investigate a large variety of neurological and neuropsychiatric disorders,
in the present study they have been employed with the purpose of detecting neuroanatomical
and gender-related abnormalities in children with autism spectrum disorders (ASD). In fact,
the dierences in the neuroanatomy of young children with ASD are an intriguing and still
poor investigated issue.
After a description of the physical principles of nuclear magnetic resonance and an overview
of magnetic resonance imaging, we specied the two algorithms that represent the object
of the current study: voxel-based morphometry and support vector machines classication
methods. Hence, we described the theoretical principles they are based on, pointing out
schemes and procedures employed to implement these analysis approaches. Then, we examined
the application of VBM and SVM methods to an opportunely chosen sample of MRI data.
A total of 152 structural MRI scans were selected. Specically, our dataset was composed
by 76 ASD children and 76 matched controls in the 2-7 year age range. The images were preprocessed applying the SPM8 algorithm, based on the dieomorphic anatomical registration
through exponentiated lie algebra (DARTEL) procedure. The resulting grey matter (GM)
segments were analyzed by applying the conventional voxel-wise two-sample t-test VBM analysis and employing the stringent family-wise error (FWE) rate correction according to random
gaussian elds theory.
The same preprocessed GM segments were then analyzed using the SVM pattern classication approach, that presents the advantage of intrinsically taking into account interregional
correlations. Moreover, this technique would allow investigations about the predictive value
of structural MRI scans. In fact, the SVM classication capability can be quantied in terms
of the area under the receiver operating characteristic curve (AUC). The leave-pair-out cross-
validation protocol has been adopted to evaluate the classication performance. The recursive
feature elimination (RFE) procedure has been implemented both to reduce the large number of features in the classication problem and to enhance the classication capability. The
SVM-RFE allows also to localize the most discriminant voxels and to visualize them in a
discrimination map. However, the pattern classication method was not employed to predict
the class membership of undiagnosed subjects, but as a gure of merit allowing to determine
an optimal threshold on the discrimination maps, where possible between-group structural
dierences are encoded.
With the aim of strengthening the SVM-based methods applied to brain data and to
guarantee reliability and reproducibility of the results, we set up the following tests:
1. We evaluated the consistency among all discrimination maps, each obtained from one
of the SVM leave-pair-out cross-validation steps, within the chosen range of number of retained features employed.
2. We assessed the dependency on the population of the training set within the cross-
validation procedure. In this way we became able to check for the stability of our
statistical results with respect to the number of subjects employed during the learning
phase. Furthermore, we can evaluate the classication performances for dierent cross-
validation schemes.
Among the results we obtained, we found that SVMs applied to GM scans correctly discriminate ASD male and female individuals with respect to controls with an AUC above the
87% with a fraction of retained voxels in the 0.4-29% range. By choosing as operative point
of the system that corresponding to the lower amount of signicant voxels (0.4% of the total
number of voxels) we obtained a sensitivity of 82% and a specicity of 80%. The resulting
discrimination maps showed some signicant regions where an excess of GM characterizes the
ASD subjects with respect to the matched control group. These regions seemed to be consistent with those obtained from the VBM analysis, nevertheless the SVM analysis highlighted
a larger number of interesting gender-specic discriminating regions.
Hence, multivariate methods based on the SVM could contribute not only to distinguish
ASD from control children, but also to disentangle the gender specicity of ASD brain alterations, consistently with respect to the mass-univariate approach.
Achieving a better AUC could make possible to employ the pattern recognition approach
not only to individuate brain regions discriminating between patients and controls, but also
to predict the class membership of undiagnosed subjects, thus facilitating the early diagnosis
of the ASD pathology.