Откривање на знаење во мозочни мрежи

Abstract

Representing brain dynamics as a complex network is standard practice in neuroscience. In order to provide functional neuroimaging data with network representation, it is generally possible to map 106 voxels from a typical fMRI protocol to a network and to study the topological properties emerging from the relationships form its nodes. However, calculating connections between large number of units can be challenging from both a computational and an interpretational point of view. The dimensionality reduction involves non-trivial renormalization of the anatomical space into functionally separate regions of interest (ROI). This process can be carried out in different ways, depending on whether it is analyzed from the anatomical principle of the spatial parcelization and the integration rules involved. Additionally, quantifying variations in the connections between different nodes in the brain across different populations and assessing their functional significance are very non-trivial tasks. Also, this advanced network overview entails a high-dimensional cost, which prevents the capability to decipher the main mechanisms behind the pathologies, and the meaning of statistical and/or machine learning tasks used to process the data. The link selection method, which allows the removal of irrelevant links in a given scenario, is a clear solution that can improve the use of these network representations. In this doctoral dissertation we look at methods of construction and analysis of brain networks from multiple views. First, wå reconstruct functional brain networks by combining different methods for links renormalization, based on frequentist and Bayesian proposed approaches and threshold strategies. Then, we asses the extent to which network topological characteristics are affected to distinguish between subjects with neurological disorders and a control group. Moreover, we show that differences in link rankings cause specific geometries in an auxiliary space that can often easily be detected with a simple visual check. We overview and a large set of statistical methods and machine learning methods for link selection and make assessment of real functional brain networks. At the end of the work, a web-based system for processing brain networks with different embedding algorithms is presented. The results of this dissertation show that different methods for link reconstruction can lead to quantitatively differences between different populations, but do not affect overall discriminatory power. Differences between populations are more visible with proportional, instead of fixed binary thresholds, while the atlas choice is not important. Link ranking can also provide a fast and reliable criterion for network reconstruction parameters for which no theoretical guidance has been proposed. From the considered statistical methods and machine learning methods for selecting links, results show that the methods act in a qualitatively similar way. While machine learning methods are conceptually more complicated than statistical ones, they do not present a clear advantage. At the same time, the high heterogeneity in the set of links kept by any method, suggests that they offer complementary dataset views. Although power to discriminate functional brain networks is not globally affected, care should be taken in the choice of available methods and in interpreting the results they give. The dissertation also discusses the implications of these results in neuroscience. Finally, the developed web-based system allows users to quickly and easily find the best method for their brain network problem and download the vector representation for further analysis.