A General Framework for the Analysis of Neuroimaging Data by Fisher Vector Descriptors and Its Use for the Improvement of Parkinson’s Disease Diagnosis (P4.111)
Neurology(2016)
摘要
Objective: present a methodology for the analysis of neuroimaging data.Background: Advanced multivariate statistical techniques for classification are widely used in neuroimaging to unveil hidden patterns in the data, allowing for modelling diseases as coordinated and causally related set of brain alterations that manifest the clinical phenotypes. However, these approaches also present a crucial shortcoming: typically small sample sizes contrast with the huge amount of information extracted per subject (dimensionality of the data). Unresolved, this hampers the validity of the methods and the results obtained. Although several techniques aim at reducing the dimensionality of the data by selecting a smaller subset of features, these adhoc approaches often discard useful information.Methods: We present a methodology for the analysis of neuroimaging data that computes descriptors by the Fisher vector algorithm. In a nutshell, the proposed approach encodes the extracted information as a measure of the deviation from a Gaussian mixture model that captures the underlying distribution of the data thereby reducing its dimensionality, and improving the linear separability of the information encoded in the description space. This methodology was applied to reanalyze diffusion tensor imaging data of the putamen in a cohort of 50 patients with early Parkinson’s disease (PD) and 50 controls. We compare our method to a more direct approach that directly feeds the raw data into the classifier.Results: By applying the proposed methodology, the classification accuracy -PD patients versus controls- was improved from 63.5[percnt] to 77.2[percnt] in comparison to the more conventional analysis.Conclusions: Our results show how, by a correct encoding of the extracted data, we can achieve a substantial improvement of the classification accuracy when using neuroimaging information and multivariate classification techniques. We propose the validation of this novel method in larger clinical cohorts and other neurodegenerative diseases. Disclosure: Dr. Salamanca has nothing to disclose. Dr. Diederich has nothing to disclose. Dr. Skupin has nothing to disclose.
更多查看译文
AI 理解论文
溯源树
样例
生成溯源树,研究论文发展脉络