Outline of the module
The module “Advanced Design & Analysis fMRI II” deals with advanced multivariate techniques to extract useful information from fMRI signal.
Standard univariate statistical models for fMRI ignore the relationships between different volume elements (voxels), by treating them separately. This relationship, however, is of great importance to understand a) how different areas of the brain co-vary and may communicate (functional connectivity) and b) how different areas jointly encode a stimulus or an experimental design. The multivariate techniques that enable overcoming these two limitations can be broadly divided into the two classes of “unsupervised” and “supervised” learning algorithms. The module will deal in detail with methods from both approaches, starting with unsupervised approaches, such as Principal Component Analysis (PCA), Independent Component Analysis (ICA) and clustering, which extract, in a data-driven fashion, sets of functionally connected brain regions from fMRI data. These methods are very useful for explore the data and separate complex artifacts from real effects.
The module will then consider supervised learning approaches, such as Multi-Voxel Pattern Analysis (MVPA) and Multivariate Regression, which are used to associate distributed patterns of activation to experimental conditions or manipulations (brain-based decoding). These methods have become very popular, as they ensure higher sensitivity to small effects compared to the standard univariate statistics.
Finally, the module will describe recently developed approaches that enable linking computational and quantitative models with patterns of fMRI activity (encoding). This is very useful for a correct and precise interpretation of fMRI experiments. Throughout the module, the participants will follow practical computer sessions where the principles and techniques covered in the module will be implemented and employed to analyse real data.
The module covers the following topics:
• Details of the available multivariate techniques for fMRI data analysis
• Application of techniques such as ICA in fMRI data analysis
• Use of pattern recognition techniques in fMRI data analysis
• What are the encoding models and how are they used in fMRI data analysis?
• Understanding of the constraints posed by the use of such techniques on experimental design
Learning objectives
At the end of this module, students will have knowledge of:
• Theoretical aspects related to multivariate modelling
• Unsupervised learning techniques (PCA, ICA, clustering) and their application to fMRI data analysis.
• Supervised learning techniques for fMRI data analysis, including MVPA and Multivariate regression
• Encoding models for fMRI data analysis.
Content
Functional MRI is employed in a large amount of studies that investigate brain function, thanks to its high spatial accuracy and its non-invasiveness.
Standard univariate analyses (GLM), however, provide only a partial view of the complexity of brain signals, ignoring valuable information, such as that contained in the co-variation of BOLD signal of different regions (functional connectivity) and that contained in distributed patterns (decoding and encoding models). Another relevant dichotomy of these multivariate data analysis tools is between unsupervised learning approaches (such as Independent Component Analysis) and supervised learning approaches (such as MVPA).
The theoretical principles behind the machine learning techniques used in fMRI, as well as their application will be discussed in the module: which approaches can be used to study correlations among different brain regions? Which principles can be used to extract spatio-temporal patterns of brain activity? How do decoding and encoding models work? Which classifiers can be used for fMRI data analysis? How should an fMRI experiment be designed, to be optimal for decoding? Which conclusions can be drawn with multivariate approaches?
This module will answer these questions in form of lectures, and in practical sessions in which the students will implement these multivariate approaches to analyse real fMRI data.
Overview of tasks and lectures
There will be 10 lectures of 2 hours distributed over 5 days.
• Introduction to multivariate models for fMRI data analysis
• Unsupervised learning: PCA and clustering
• Unsupervised learning: ICA, general principles
• Unsupervised learning: ICA, applications to fMRI data
• Supervised Learning: introduction to MVPA
• Supervised Learning: applications of MVPA to fMRI data
• Supervised Learning: introduction to Multivariate Regression
• Supervised Learning: application of Multivariate Regression to fMRI data.
• Supervised Learning: introduction to Encoding models
• Supervised Learning: application of Encoding models to fMRI.
Position within the programme
This is a unique module in this Master programme dealing with advance multivariate techniques for the design principles and analyses steps employed in fMRI studies. The knowledge of these analysis tools is highly relevant for a correct design of fMRI data and for a correct interpretation of the results. This module is complementary to the general modules dealing with MRI physics and data analysis basics and to the modules dealing with advanced analysis methods and applications.
Teaching format
Structure
The module is a one week-long residential module consisting of 10 lectures of 2 hours. In addition, the students will have computer sessions where, under the guidance of tutors, they will implement the analyses approaches discussed during the lectures and analyse fMRI data. Furthermore, the residential part is combined with a preparatory reading phase and post-module marked assignments.
Grading
Passing the module requires an 85% attendance to the lectures and practical sessions, and a satisfactory completion of the practical sessions and the module assignments. The module assignments will be summarised by the students in a written form which will be evaluated by the module coordinator(s).