A Group Approach to Functional Cortical Parcellation from Resting-State fMRI

3778 

Abstract Submission 

Wednesday, June 17 & Thursday, June 18 

Minqi Chong¹, Anand Joshi¹, Justin Haldar¹, Elizabeth DuPre², Wen-Ming Luh², David Shattuck³, R. Nathan Spreng², Richard Leahy¹

¹University of Southern California, Los Angeles, CA, ²Cornell University, Ithaca, NY, ³University of California, Los Angeles, Los Angeles, CA

Resting-state fMRI can be used to generate functional parcellations of cerebral cortex and the correlations between these parcels then used to infer functional networks. Using functional rather than purely anatomical features to drive parcellation produces correspondences across groups that can lead to superior power in investigating functional specialization and networks. We describe a group approach to simultaneous parcellation and network identification through a combination of graph-cut based image segmentation and inverse covariance estimation using group-sparsity. This approach differs from alternative approaches [1,2] through combining group-sparsity based network identification with parcellation.

We assume that a functional parcellation of cerebral cortex exists which has the same number of regions in each subject. These regions form the nodes of a graph and the connections between nodes, as represented by non-zero partial correlation values, is the same for all subjects, although the strength of these connections can vary. For each subject we identify the boundaries of the regions based on a Markov random field model [3] constrained by a data fitting term which consists of a multivariate Gaussian (MVG) likelihood for each surface element that computes the likelihood of each element on the cortical surface belonging to each region. The covariances of the MVG model are themselves regularized using a group sparsity penalty [4] that encourages each subject to have the same set of non-zero elements in the inverse covariances between the time series as averaged over each region. We alternate between optimization with respect to the labeling of each subject and estimation of their partial correlation matrices. The former is performed using a normalized graph-cut algorithm [5], the latter using an ADMM approach [1]. Convergence was observed empirically for synthetic and in vivo data.
Synthetic data consisted of 32 separate cortices with parcellations and rs-fMRI time series synthesized to reflect the above model assumptions. In vivo resting state fMRI time series data were acquired together with structural MRIs for 32 subjects. The multiecho fMRI data were pre-processed and de-noised in AFNI [6]. Cortical surfaces were generated and structurally aligned to an atlas using BrainSuite. The atlas contained 80 anatomical ROIs across two cortical hemispheres including subparcellations of ROIs to include regions within the default network (DN) [7]. We show results below for a subset of 4 subjects, jointly processed as described above.

Fig. 1 illustrates convergence of the alternating optimization procedure both in terms of labeling of cortical ROIs and estimation of the associated partial correlation networks. Time to convergence will increase as the degree of required modification of the ROIs increases.
We show the evolution of the ROIs in the DN in Fig. 2 for four subjects as a function of iteration. In these images, the subject cortices are mapped back to the atlas surface using the fixed anatomical mapping computed with BrainSuite. Visually comparing iterative changes, we see regions growing or shrinking across subjects. As a measure of performance we also look at the following measure of uniformity within each ROI vs. iteration in Fig. 3. For each ROI we compute the ratio of squared first singular value to total energy in the spatio-temporal data matrix. As the time series across each individual ROI become more similar, a larger fraction of energy should be represented in the first singular value and this measure should increase, as it does in Fig. 3.

Our group parcellation approach with group sparse inverse covariances and a Gaussian Markov random field model converges to plausible parcellations that differ from their anatomically-driven counterparts. Our results also show the resulting parcellations produce improved uniformity of functional signals across ROIs.

Classification and Predictive Modeling

fMRI Connectivity and Network Modeling ¹

Image Registration and Computational Anatomy

Task-Independent and Resting-State Analysis ²

Cortical Anatomy and Brain Mapping

Poster Session - Wednesday

Data Registration

FUNCTIONAL MRI

Segmentation