In this study, we propose a hierarchical network representation of human brain extracted from fMRI data. This representation consists of two levels. In the first level, we form a network among the voxels, smallest building block of fMRI data. In the second level, we define a set of supervoxels by partitioning the first level network into a set of subgraphs, which are assumed to represent homogeneous brain regions with respect to a pre-defined criteria. For this purpose, we develop a novel brain parcellation algorithm, called BrainParcel. As current literature tends to represent human brain as a graph, BrainParcel adopts this approach. The suggested algorithm partitions a brain network, called mesh network using a graph partitioning method. The supervoxels obtained at the output of BrainParcel form partitions of brain as an alternative to anatomical regions. Compared to anatomical regions, supervoxels gather the linearly dependent voxels. As the next step, we form a mesh network among the supervoxels. Therefore, we represent fMRI data by two networks of different granularity. The first network is at voxel level, whereas the second network is at supervoxel level.
In order to test the representation power of this two level network, we suggest an ensemble learning architecture, called Cognitive Learner. The suggested ensemble learning method is used in brain decoding problem, where we classified the cognitive states. The results applied on an object recognition problem shows that the suggested BrainParcel algorithm together with Cognitive Learner has a better representation power on brain decoding in terms of classification accuracy.