- S. Allassonnière, CMAP, Polytechnique
- N. Paragios, ECP
Human brain project partners:
- Julich research, center, Germany
- UCL, London
Mapping brain functional connectivity from functional Magnetic Resonance Imaging (MRI) data
has become a very active field of research. However, analysis tools are limited and many impor-
tant tasks, such as the empirical definition of brain networks, remain difficult due to the lack of a
good framework for the statistical modeling of these networks. We propose to develop population
models of anatomical and functional connectivity data to improve the alignment of subjects brain
structures of interest while inferring an average template of these structures. Based on this es-
sential contribution, we will design new statistical inference procedures to compare the functional
connections between conditions or populations and improve the sensitivity of connectivity analy-
sis performed on noisy data. Finally, we will test and validate the methods on multiple datasets
and distribute them to the brain imaging community.
Introduction: In human brain imaging, the inaccuracy of cross-subject registration is a well-identified bottleneck for data analysis. Registration is an important field of research in which many groups are investing resources; these efforts are almost entirely based on the information carried by anatomical images. Yet, there is rich and relevant information in functional connectivity data that might be used to improve cross-subject registration in regions where the anatomical information is poorly informative or unreliable. Such an effect has been shown for functional data 1 and is expected to play a major impact in connectome mapping. Our proposition to revisit cross-subject registration by taking into account functional connectivity data addresses a central technical issue for multi-subject connectome mapping.
Brain image registration: Most modern registration algorithms (see 2 for a user-oriented review, 3 for a technical review) impose diffeomorphic registration models to warp a given image to a template by matching some contrast, and rely on alternate template and warping estimation. While the use of functional connectivity information is a priori well suited to inform the correspondence between subjects 1, it is still unclear how a computationally tractable representation of such data can be constructed, given their lack of salient features. Our hope is that Dictionary Learning 4 will provide potentially useful contrasts to match multisubject datasets, as these analytic tools have a denoising effect: i) through the regularizing prior commonly used in these settings ii) and by enforcing cross-subject consistency that tends to regularize the information observed in some subjects based on group-level information.
The project consists in blending functional information into brain image registration algorithms. We will inject functional information by coupling a deformation framework with a set of contrasts, that comprise spatial maps obtained from multi-subject dictionary learning procedures 4, in order to jointly estimate functional regions and coregister individual data. Regarding the registration problem, we will consider different standard alternative models, the cost and merit of which (in terms of cost and accuracy) will be assessed carefully:
• The diffeomorphic log-daemons framework, which is efficient, has proven to be effective in many contexts and is well-mastered by our lab (PhD thesis of V. Siless).
• Several avatars of the LDDMM framework 5, that is considered as the state-of-the art approach
• Discrete optimization approaches 3 for multi-modal cross-subject registrations, that has particular strengths, such as a lesser sensitivity to initialization.
Our aim is to provide public open-source implementations of the methods developed as part of this thesis. The functions developed to solve functional data analysis problems, after a thorough assessment, will be made available with a homogeneous Application Program Interface (API); these libraries, in particular Nilearn http://nilearn.github.io or Nipy http://nipy.org/nipy, rely as much as possible on existing functions of the Python ecosystem.