Analisi di immagini biomediche per la ricerca di biomarcatori del Disturbo dello Spettro Autistico
Year: 2017
Type: dissertation
Abstract: [English] In this thesis, the problem of the research of a neuroanatomical biomarker for Autism with Nuclear Magnetic Resonance (NMR) brain images has been studied. Despite the efforts put and the various approaches tried, this kind of studies still has not reported a meaningful and shared result. I have therefore examined the methodologies of analysis of NMR data based on Machine Learning techniques, both in their basic assumptions and in their implementation. Usually, these analyses are structured in three steps: (i) image gathering and selection, (ii) brain segmentation, extraction of neuroanatomical parameters and their normalization, (iii) use of Machine Learning algorithms to detect a recognizing pattern between data of healthy subjects and those with Autism. However there is not a shared agreement regarding how these steps are actually implemented. In fact often each study follows different rules, often without verifying the effectiveness and correctness of the approach chosen. It emerged that the most common practices in this sector are based on wrong assumptions and can therefore influence the result of the study. In particular the analyses conducted showed that: • The typical criterion of data selection, based on visual inspection of NMR data is insufficient, since it is not able to assess the goodness of the anatomical quantities extracted after the brain segmentation. In fact it has been verified that 16% of subjects that after the segmentation process had anomalous values actually had good NMR images. • the quantities extracted with the most used segmentation tool in literature, Freesurfer, are characterized by a strong dependence on the NMR scanner and its parameters set for the acquisition. In fact, Machine Learning algorithms recognize with subjects acquired with different scanner parameters better than males and females. This finding severily questions the validity of multi-site studies. • The most common normalization method of brain quantities, usually done in order to eliminate the dependency of the data from age and skull dimensions, is not effective and therefore only adds noise to the data. After having identified the incorrect procedures and defined an adequate analysis methodology following the principles of Knowledge Discovery in Database (KDD), the performances of the classification of healthy and Autistic subjects have been assessed. This analysis has been done first using the current definition of the disease of Autism Spectrum Disorder (ASD), described in the fifth edition of the Diagnostic Statistical Manual of mental diseases (DSM). Then, the same analysis has been repeated dividing the subjects in the diagnostic categories identified in the earlier DSM version, in which ASD subjects were divided in three more homogeneous classes: Autism Disease (AD), Asperger Syndrome (AS) and Pervasive Developmental Disorder, Not Otherwise Specified (PDD-NOS). The classifier trained to recognize healthy subjects from ASDs reaches a modest performance of 59% (measured with AUC: Area Under the Receiver Operating Characteristic Curve). If instead the diagnostic categories of AD and PDD-NOS are used, performances go up to 63% and 82% AUC respectively. The subjects affected by Asperger Syndrome have not been found to be statistically distinguishable from healthy individuals. In conclusion, in this thesis the existence of numerous sources of errors has been shown in the most common procedures used for neuorimages analysis. These errors may have contributed the inconsistency between the results available in literature. Furthermore, based on the new research methodology described in this work, it has been shown that the etherogeneity of the disease have a detrimental effect on the analysis. The obtained results, are a quantitatively consolidated starting point to direct future studies, that must be oriented towards the investigation of new diagnostic categories, better characterizable from the neuroanatomical point of view. [Italiano]
Author: FERRARI, ELISA
Institution: University of Pisa
Language: English
Subfield: Cognitive Neuroscience
Field: Neuroscience
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W6979855885
