Statistical Analysis of NeuroImaging Data

Functional magnetic resonance imaging (fMRI) is a noninvasive tool for studying brain activity. This project validates and improves existing statistical models for fMRI data. The results show that the statistical methods traditionally used in analyzing fMRI can result in a high level of false positives. The results were published in the scientific journal PNAS and created quite a lot of media attention. 

Functional magnetic resonance imaging (fMRI) is a popular tool for studying brain activity. It can non-invasively image the human brain without any ionizing radiation. The method can detect brain activity in correlation to tasks performed by the subject during the scan and is often used in neurology and psychology research.

Complex Noise

From a statistical perspective, analyzing fMRI data is a challenging task for several reasons. One reason is that the noise created during the scan has a complex structure, which is virtually impossible to simulate in a computer. Another reason is that there are several noise sources which distort the signal of interest, for example head motion, breathing and pulse. Also, the MRI itself creates a variation of electronic noise.

-If you move your head it affects the magnetic field and that effect lingers for a while and does not only disturb the image in that moment but also the following images, says Associate Professor Anders Eklund.

Anders has one foot in biomedical engineering and the other in computer science with experience in fMRI method development, statistics and image analysis.

Improving Existing Models

This project validates and improves existing statistical models for neuroimaging data. Most fMRI research projects use the method as a tool and does not reflect on the method itself.

-I think that too few researchers are studying the actual method and how to analyze the data. If everyone is analyzing their data with the same method and that method later on turns out to be wrong. Then a lot of studies will be affected.

Thinking about this Anders became more and more interested in how to analyze the data and how you can verify what is correct and what is not. This type of research requires large datasets where you can run multiple analyses on normal subjects.

-We use open data available through data sharing. Even though we have great opportunities to scan and collect our own material at CMIV, it is very time consuming and expensive to scan 1000 subjects.

In this project Anders is collaborating with Mattias Villani who is a Professor in Bayesian statistics at Linköping University and Stockholm University. His background is in econometrics, analyzing economic time series.

-In fMRI you have time series as well so we could apply most of the methods directly. After that the methods grew more and more advanced, Anders explains.

Another key player in the project is Professor Thomas Nichols from University of Oxford. He has long experience in neuroimaging statistics and became an important sounding board.

-I had plenty of time during my PostDoc for running all the analyses. Tom had a more pressed schedule but a lot of knowledge to contribute with. Despite communicating almost exclusively by email it turned out to be a perfect collaboration, says Anders.

Media Attention

Anders and his colleagues found that the statistical methods traditionally used in analyzing fMRI can result in a high level of false positives. The results were published in the scientific journal PNAS and created quite a lot of media attention.

In a follow-up study the research group extended their work and responded to some of the questions and criticism the first article received.

-We previously showed that the non-parametric permutation test can perform better than commonly used parametric tests, as the permutation test is based on a lower number of statistical assumptions.

However, in a few cases even the permutation test produced invalid results.

-We investigated several ways to obtain nominal false positive rates, and finally discovered that physiological noise can disturb the group analyzes. To correctly model physiological noise requires monitoring of breathing and pulse during the fMRI experiment, Anders continues.

After eleven years of research on fMRI Anders is expanding his research into deep learning.

-I think that it is important not to get stuck in the same tracks for too long. The things I learn in my deep learning projects can give me inspiration for the fMRI research as well. That’s also why my collaboration with Mattias Villani is so important. He brings other angles, coming from another field.

 

Key Publications
Show/Hide content

Anders Eklund, Martin A Lindqvist, Mattias Villani (2017)

NeuroImage , Vol.155 , s.354-369 Continue to DOI

Anders Eklund, Thomas Nichols, Hans Knutsson (2016)

Proceedings of the National Academy of Sciences of the United States of America , Vol.113 , s.7900-7905 Continue to DOI

Anders Eklund, Hans Knutsson, Thomas E Nichols (2019)

Human Brain Mapping , Vol.40 , s.2017-2032 Continue to DOI

Contacts
Show/Hide content