Publicação
The effect of using multiple connectivity metrics in brain Functional Connectivity studies
| Resumo: | Resting-state functional magnetic resonance imaging (rs-fMRI) has the potential to assist as a diagnostic or prognostic tool for a diverse set of neurological and neuropsychiatric disorders, which are often difficult to differentiate. fMRI focuses on the study of the brain functional Connectome, which is characterized by the functional connections and neuronal activity among different brain regions, also interpreted as communications between pairs of regions. This Functional Connectivity (FC) is quantified through the statistical dependences between brain regions’ blood-oxygen-level-dependent (BOLD) signals time-series, being traditionally evaluated by correlation coefficient metrics and represented as FC matrices. However, several studies underlined limitations regarding the use of correlation metrics to fully capture information from these signals, leading investigators towards different statistical metrics that would fill those shortcomings. Recently, investigators have turned their attention to Deep Learning (DL) models, outperforming traditional Machine Learning (ML) techniques due to their ability to automatically extract relevant information from high-dimensional data, like FC data, using these models with rs-fMRI data to improve diagnostic predictions, as well as to understand pathological patterns in functional Connectome, that can lead to the discovery of new biomarkers. In spite of very encouraging performances, the black-box nature of DL algorithms makes difficult to know which input information led the model to a certain prediction, restricting its use in clinical settings. The objective of this dissertation is to exploit the power of DL models, understanding how FC matrices created from different statistical metrics can provide information about the brain FC, beyond the conventionally used correlation family. Two publicly available datasets where studied, the ABIDE I dataset, composed by healthy and autism spectrum disease (ASD) individuals, and the ADHD-200 dataset, with typically developed controls and individuals with attention-deficit/hyperactive disorder (ADHD). The computation of the FC matrices of both datasets, using different statistical metrics, was performed in MATLAB using MULAN’s toolbox functions, encompassing the correlation coefficient, non-linear correlation coefficient, mutual information, coherence and transfer entropy. The classification of FC data was performed using two DL models, the improved ConnectomeCNN model and the innovative ConnectomeCNN-Autoencoder model. Moreover, another goal is to study the effect of a multi-metric approach in classification performances, combining multiple FC matrices computed from the different statistical metrics used, as well as to study the use of Explainable Artificial Intelligence (XAI) techniques, namely Layer-wise Relevance Propagation method (LRP), to surpass the black-box problem of DL models used, in order to reveal the most important brain regions in ADHD. The results show that the use of other statistical metrics to compute FC matrices can be a useful complement to the traditional correlation metric methods for the classification between healthy subjects and subjects diagnosed with ADHD and ASD. Namely, non-linear metrics like h2 and mutual information, achieved similar and, in some cases, even slightly better performances than correlation methods. The use of FC multi-metric, despite not showing improvements in classification performance compared to the best individual method, presented promising results, namely the ability of this approach to select the best features from all the FC matrices combined, achieving a similar performance in relation to the best individual metric in each of the evaluation measures of the model, leading to a more complete classification. The LRP analysis applied to ADHD-200 dataset proved to be promising, identifying brain regions related to the pathophysiology of ADHD, which are in broad accordance with FC and structural study’s findings. |
|---|---|
| Autores principais: | Teixeira, Hugo Emanuel Augusto |
| Assunto: | Conetividade Funcional Conectoma Distúrbios Cerebrais Redes Neuronais Profundas Explicabilidade de Redes Neuronais Teses de mestrado - 2022 |
| Ano: | 2022 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade de Lisboa |
| Idioma: | inglês |
| Origem: | Repositório da Universidade de Lisboa |
| Resumo: | Resting-state functional magnetic resonance imaging (rs-fMRI) has the potential to assist as a diagnostic or prognostic tool for a diverse set of neurological and neuropsychiatric disorders, which are often difficult to differentiate. fMRI focuses on the study of the brain functional Connectome, which is characterized by the functional connections and neuronal activity among different brain regions, also interpreted as communications between pairs of regions. This Functional Connectivity (FC) is quantified through the statistical dependences between brain regions’ blood-oxygen-level-dependent (BOLD) signals time-series, being traditionally evaluated by correlation coefficient metrics and represented as FC matrices. However, several studies underlined limitations regarding the use of correlation metrics to fully capture information from these signals, leading investigators towards different statistical metrics that would fill those shortcomings. Recently, investigators have turned their attention to Deep Learning (DL) models, outperforming traditional Machine Learning (ML) techniques due to their ability to automatically extract relevant information from high-dimensional data, like FC data, using these models with rs-fMRI data to improve diagnostic predictions, as well as to understand pathological patterns in functional Connectome, that can lead to the discovery of new biomarkers. In spite of very encouraging performances, the black-box nature of DL algorithms makes difficult to know which input information led the model to a certain prediction, restricting its use in clinical settings. The objective of this dissertation is to exploit the power of DL models, understanding how FC matrices created from different statistical metrics can provide information about the brain FC, beyond the conventionally used correlation family. Two publicly available datasets where studied, the ABIDE I dataset, composed by healthy and autism spectrum disease (ASD) individuals, and the ADHD-200 dataset, with typically developed controls and individuals with attention-deficit/hyperactive disorder (ADHD). The computation of the FC matrices of both datasets, using different statistical metrics, was performed in MATLAB using MULAN’s toolbox functions, encompassing the correlation coefficient, non-linear correlation coefficient, mutual information, coherence and transfer entropy. The classification of FC data was performed using two DL models, the improved ConnectomeCNN model and the innovative ConnectomeCNN-Autoencoder model. Moreover, another goal is to study the effect of a multi-metric approach in classification performances, combining multiple FC matrices computed from the different statistical metrics used, as well as to study the use of Explainable Artificial Intelligence (XAI) techniques, namely Layer-wise Relevance Propagation method (LRP), to surpass the black-box problem of DL models used, in order to reveal the most important brain regions in ADHD. The results show that the use of other statistical metrics to compute FC matrices can be a useful complement to the traditional correlation metric methods for the classification between healthy subjects and subjects diagnosed with ADHD and ASD. Namely, non-linear metrics like h2 and mutual information, achieved similar and, in some cases, even slightly better performances than correlation methods. The use of FC multi-metric, despite not showing improvements in classification performance compared to the best individual method, presented promising results, namely the ability of this approach to select the best features from all the FC matrices combined, achieving a similar performance in relation to the best individual metric in each of the evaluation measures of the model, leading to a more complete classification. The LRP analysis applied to ADHD-200 dataset proved to be promising, identifying brain regions related to the pathophysiology of ADHD, which are in broad accordance with FC and structural study’s findings. |
|---|