Multi-subject Dictionary Learning to Segment an Atlas of Brain Spontaneous Activity

Abstract

Thirion. Multisubject dictionary learning to segment an atlas of brain spontaneous activity. Information Processing in Medical Imaging, Jul 2011, Kaufbeuren, Germany. Springer, 6801, pp.562-573, 2011, Lecture Notes in Computer Science. <10.1007 Multi-subject dictionary learning to segment an atlas of brain spontaneous activity Abstract. Fluctuations in brain on-going activity can be used to reveal its intrinsic functional organization. To mine this information, we give a new hierarchical probabilistic model for brain activity patterns that does not require an experimental design to be specified. We estimate this model in the dictionary learning framework, learning simultaneously latent spatial maps and the corresponding brain activity time-series. Unlike previous dictionary learning frameworks, we introduce an explicit difference between subject-level spatial maps and their corresponding population-level maps, forming an atlas. We give a novel algorithm using convex optimization techniques to solve efficiently this problem with non-smooth penalties well-suited to image denoising. We show on simulated data that it can recover population-level maps as well as subject specificities. On resting-state fMRI data, we extract the first atlas of spontaneous brain activity and show how it defines a subject-specific functional parcellation of the brain in localized regions.

Cohort-Level Brain Mapping: Learning Cognitive Atoms to Single Out Specialized Regions

“…The algorithm iterates over small batches of voxels (hundreds) to incrementally improve the dictionary. When the number of voxels is large, such an approach can be orders of magnitude faster than the alternate optimization strategies used by [18,3], because these require solving brain-wide sparse regression for each update of the dictionary. Szabo et al [14] extend this approach to structured dictionaries by replacing the 1 norm on α v with a structure-inducing norm, such as the 21 norm used in the group lasso.…”

“…Laird et al [6] have recently shown that the modes that it extracts from task-activation data capture meaningful structure in the space of cognitive processes. Beyond ICA, Varoquaux et al [18] use dictionary learning to segment a functional parcellation from resting-state. Very interesting preliminary work by Chen et al [3] integrates spatial normalization with dictionary learning to estimate jointly an inter-subject warping and functional regions.…”

Statistical Learning for Resting-State fMRI: Successes and Challenges

“…While multi-subject analyses do stabilize model estimation (see Fig. 1), at very high model order the spatial maps may reflect modes of inter-subject variability, such as spatial gradients of some of the networks [19].…”

“…Combined with PCA in a sparse PCA procedure, it regularizes model estimation by constraining the fitted signal [19]. As ℓ 1 penalizations perform poorly with correlated features, it is useful to impose sparsity on groups of neighboring voxels, for instance using sparse structured penalties.…”

“…sharing the same spatial maps V. However, for ICA-based approaches, separating signal-space selection at the subject and the group level stabilizes estimation in high-model-order settings and extracts more meaningful maps [18]. In a sparse decomposition framework, group-level maps can be formulated as a prior on subject-level models, as in the Multi-Subject Dictionary Learning (MSDL) approach [19]. The corresponding joint estimation of subjectlevel and group-level maps yields a reference population atlas of brain networks and corresponding subject-specific counter-parts.…”