Computational analysis of schizophrenia: Implementation of a multivariate model of anatomical differences

Abstract

[eng] Chronic schizophrenia has been widely studied, consistent findings have shown the anatomical pattern associated with this disease, but the clinical picture is often undifferentiated at first presentation. Finding morphometric alterations associated with a disease is a widespread goal in neuroimaging. It has been performed in hundreds of studies from applying Voxel-Based Morphometry (VBM). However, VBM is a mass-univariate approach, assumes that voxels are independent and this may not be the most biologically plausible assumption to make. Many neuroimaging advances are focused in multivariate framework, like pattern recognition approaches. Such applications reveal complex associations and prediction models that provide greater accuracy for characterizing differences, also in schizophrenia. Such techniques require some form of characterization of inter-subject neuroanatomical variability, where the registration plays a significant role. If data are imprecisely modeled or the characterization used does not incorporate key information, this may result in poor predictions. The use of suboptimal features limits the accuracy with which predictions may be made. This scenario makes necessary exploring features from images and use the most informative to optimise pattern recognition in clinical research. Regarding modeling, accuracy is being established during VBM-type preprocessing. If segmentation does not work accurately, the next normalisation step cannot be accurate either. Hence the interest in the accuracy of automated computational tools is also increasing. To adress these issues, the current thesis was divided into three studies. First study was focused on the comparison between segmentation algorithms by SPM (http://www.fil.ion.ucl.ac.uk/spm/): “Unified segmentation” (US) and “New Segmentation” (NS), and FSL (http://fsl.fmrib.ox.ac.uk/fsl/): “FMRIB’s Automated Segmentation tool” (FAST). The IBSR dataset (http://vivo.cornell.edu/display/individual5017) that includes segmented classes by experts was used to establish a ground truth. A detailed comparison between algorithms was conducted using different methods. In study 2 a Gaussian Process machine learning approach was used for predicting age, gender and body mass index (BMI) using the IXI dataset (http://biomedic.doc.ic.ac.uk/brain-development/index.php?n=Main.Datasets). MRI data were segmented using NS and registered with the “Shooting Geodesic toolbox”. Proper characterizations from VBM-type preprocessed data (linear kernels) and its dependence on the smoothing (FWHM from 0 to 20mm) were evaluated. Study 3 consisted in an application to Schizophrenia (Sample involved 111 patients and 111 controls provided by FIDMAG: http://www.fidmag.com/fidmag/index.php) with the optimal features from study 2. Our hypothesis was that image features that worked well in study 2 would also work well for predicting schizophrenia. Results from study 1 showed that US was the most sensitive algorithm, and FAST the most specific, NS was found the most balanced of the three, no significant differences w.r.t. the sensitivity of US and the specificity of FAST were detected. Moreover, NS obtained the highest Jaccard coefficient, becoming the most similar to the ground truth. FAST was found the last in this ranking. In study 2, results from predicting age, gender and BMI pointed that scalar momentum was the best feature. Interestingly, grey matter (GM) was not the best feature for predicting age, and whithe matter was the best feature for predicting BMI. In general, performances were highly dependent on the smoothing, although scalar momentum was not among the most dependent. Findings from study 3 showed that scalar momentum provided best feature than GM for predicting schizophrenia, this results confirmed the hypothesis a priori. Main conclusion is that multivariate pattern recognition analyses using scalar momentum provide an excellent strategy for classifying schizophrenia. This approach might potentially be extended to other psychiatric and neurodegenerative diseases both in research and as an aid to differential diagnosis in routine clinical practice.