Skip to Main content Skip to Navigation

Feature extraction and supervised learning on fMRI : from practice to theory

Abstract : Until the advent of non-invasive neuroimaging modalities the knowledge of the human brain came from the study of its lesions, post-mortem analyses and invasive experimentations. Nowadays, modern imaging techniques such as fMRI are revealing several aspects of the human brain with progressively high spatio-temporal resolution. However, in order to answer increasingly complex neuroscientific questions the technical improvements in acquisition must be matched with novel data analysis methods. In this thesis we examine different applications of machine learning to the processing of fMRI data. We propose novel extensions and investigate the theoretical properties of different models. % The goal of an fMRI experiments is to answer a neuroscientific question. However, it is usually not possible to perform hypothesis testing directly on the data output by the fMRI scanner. Instead, fMRI data enters a processing pipeline in which it suffers several transformations before conclusions are drawn. Often the data acquired through the fMRI scanner follows a feature extraction step in which time-independent activation coefficients are extracted from the fMRI signal. The first contribution of this thesis is the introduction a model named Rank-1 GLM (R1-GLM) for the joint estimation of time-independent activation coefficients and the hemodynamic response function (HRF). We quantify the improvement of this approach with respect to existing procedures on different fMRI datasets. The second part of this thesis is devoted to the problem of fMRI-based decoding, i.e., the task of predicting some information about the stimuli from brain activation maps. From a statistical standpoint, this problem is challenging due to the high dimensionality of the data, often thousands of variables, while the number of images available for training is small, typically a few hundreds. We examine the case in which the target variable consist of discretely ordered values. The second contribution of this thesis is to propose the following two metrics to assess the performance of a decoding model: the absolute error and pairwise disagreement. We describe several models that optimize a convex surrogate of these loss functions and examine their performance on different fMRI datasets. Motivated by the success of some ordinal regression models for the task of fMRI-based decoding, we turn to study some theoretical properties of these methods. The property that we investigate is known as consistency or Fisher consistency and relates the minimization of a loss to the minimization of its surrogate. The third, and most theoretical, contribution of this thesis is to examine the consistency properties of a rich family of surrogate loss functions that are used in the context of ordinal regression. We give sufficient conditions for the consistency of the surrogate loss functions considered. This allows us to give theoretical reasons for some empirically observed differences in performance between surrogates.
Document type :
Complete list of metadata

Cited literature [255 references]  Display  Hide  Download
Contributor : Abes Star :  Contact
Submitted on : Tuesday, January 26, 2016 - 11:12:24 AM
Last modification on : Thursday, December 10, 2020 - 11:03:52 AM


Version validated by the jury (STAR)


  • HAL Id : tel-01100921, version 2


Fabian Pedregosa-Izquierdo. Feature extraction and supervised learning on fMRI : from practice to theory. Medical Imaging. Université Pierre et Marie Curie - Paris VI, 2015. English. ⟨NNT : 2015PA066015⟩. ⟨tel-01100921v2⟩



Record views


Files downloads