Brain parcellation

Neurologists have quickly established that the brain is organized into functional units collaborating towards performing complex tasks. The development of functional MRI (fMRI) has confirmed this fact by offering us a mean to observe the complex interactions between brain regions in vivo, almost in real time, and at a millimetric resolution. Researchers are now attempting to establish how the network of connections between the functional brain units, the connectome, develops during childhood, is altered by diseases, and declines with aging. All these analyses rely on the crucial assumption that the functional units underlying the brain have been correctly segmented. This hypothesis has been backed by developing algorithms generating high-quality data-driven functional parcellation closely fitting their input data. Multiple functional parcellation methods have been proposed through the years. I joined this effort by developing a method based on a discrete Markov Random Field framework, GraSP, specifically designed to delineate strongly coherent brain parcels.


associated publications

See software page for more details.

Subject-specific Parcellations

Brain parcellations obtained for large populations constitute very useful atlases for clinical studies. However, we know that brain organization slightly differs from subject to subject, and this variability is not taken into account when a unique atlas is used. A lot of effort has been recently dedicated to the generation of individual brain parcellations, with the hope of deriving subject-specific biomarkers useful for diagnosis/prognosis. However, the significant amount of noise corrupting the fMRI scans is a strong challenge.
During this project, I reduced the noise in individual scans and I controlled the difference between individual and population parcellations by mixing individual and population correlations. I am now investigating fMRI denoising and alternative connectivity measures with the hope of improving enough the reliability of individual parcellation to abandon this strategy (see the last section of this page for more details).

associated publications

Parcellation of the brain according to structural connectivity

A research topic is closely related to the parcellation of the brain according to functional connectivity: the parcellation of the brain according to structural connectivity, determined by running tractography algorithms on diffusion-weighted imaging scans (DTI scans, and their successors: HARDI, multi-shell acquisitions). In practice, the algorithms used for functional parcellation can be directly used for structural parcellation, after replacing the measure of functional connectivity (in my case: Pearson correlation between functional time series) by a measure of structural connectivity.
Structural connectivity is often compressed, cleaned, and made more intelligible by grouping structural tracts into large bundles of fibers connecting well-known brain regions, such as all the fibers leaving the corpus callosum, all the fibers reaching the motor cortex... The connectivity of each voxel at the interface between gray and white matter is then summarized into a short connectivity signature. I proposed a novel approach to compare these signatures when parcellating the cortex.

Cortical parcellations based on structural connectivity.

associated publication

Related questions

My research on parcellation led me to investigate the related topics listed below.
Parcellations comparisons
Comparing parcellations is not straightforward, in particular when their number of parcels differ. I have adopted several methods designed for comparing clusterings, such as the adjusted Rand index. However, these indices can only be used to compare two parcellations. When multiple parcellations are generated, it is necessary to compare them two by two, and the computational burden increases quadratically with the number of parcellations to compare. I have therefore investigated different approaches for speeding the computations when a large number of parcellations need to be compared.

associated publication

Impact of local cortical geometry
The strength and the signal-to-noise ratio of the fMRI signal measured inside the cortex are likely to be impacted by the local cortical thickness and the local folding of the brain. This statement implies that functional parcellations might be confounded by local brain geometry. I investigated this concern by researching a statistical link between the local geometric properties of the cortex, such as local curvature, and cortical thickness, and the local functional homogeneity (FH), which indicates if the BOLD signal is locally coherent. The first set of results obtained for the Philadelphia Neurodevelopment Cohort and later replicated using data provided by the Human Connectome Project pointed several very significant relations between local geometry and local functional homogeneity, which would deserve to be further investigated.
These results underline the needs for multimodal analyses. I have started to explore this direction, by running preliminary experiments involving other cortical measures such as cortical thickness and myelin, and by proposing a strategy for fusing multimodal cortical data (to be submitted soon).

Cortical myelin (left hemisphere, median over the hundred unrelated subjects of the Human Connectome Project).

associated publications

Better functional connectivity measures for better parcellations?
The most standard measure of functional connectivity is the Pearson correlation between fMRI time series. Several statistical methods improving the quality of Pearson correlations have been recently published. Also, several strategies could be implemented for denoising fMRI signals, and several alternative connectivity measures exist. This vivid literature raises the hope of addressing a well-known issue in our field: the large amount of noise in the fMRI scans, which has prevented so far the definition of robust functional biomarkers for individual diagnosis/prognosis. Due to the rich literature, at the interface between statistics, signal processing and fMRI research and development, my investigations on this topic quickly formed a project on its own, presented here.