Mapping Informative Clusters in a Hierarchial Framework of fMRI Multivariate Analysis

Citation: Xu R, Zhen Z, Liu J ( Mapping Informative Clusters in a Hierarchial Framework of fMRI Multivariate Analysis Rui Xu 0 Zonglei Zhen 0 Jia Liu 0 Olaf Sporns, Indiana University, United States of America 0 1 College of Life Science, Graduate University of Chinese Academy of Sciences , Beijing , People's Republic of China, 2 State Key Laboratory of Cognitive Neuroscience and Learning, Beijing Normal University , Beijing , People's Republic of China Pattern recognition methods have become increasingly popular in fMRI data analysis, which are powerful in discriminating between multi-voxel patterns of brain activities associated with different mental states. However, when they are used in functional brain mapping, the location of discriminative voxels varies significantly, raising difficulties in interpreting the locus of the effect. Here we proposed a hierarchical framework of multivariate approach that maps informative clusters rather than voxels to achieve reliable functional brain mapping without compromising the discriminative power. In particular, we first searched for local homogeneous clusters that consisted of voxels with similar response profiles. Then, a multi-voxel classifier was built for each cluster to extract discriminative information from the multi-voxel patterns. Finally, through multivariate ranking, outputs from the classifiers were served as a multi-cluster pattern to identify informative clusters by examining interactions among clusters. Results from both simulated and real fMRI data demonstrated that this hierarchical approach showed better performance in the robustness of functional brain mapping than traditional voxelbased multivariate methods. In addition, the mapped clusters were highly overlapped for two perceptually equivalent object categories, further confirming the validity of our approach. In short, the hierarchical framework of multivariate approach is suitable for both pattern classification and brain mapping in fMRI studies. - Funding: This study is funded by the 100 Talents Program of the Chinese Academy of Sciences, the National Basic Research Program of China (2010CB833903, 2011CB505402), the National Natural Science Foundation of China (30800295), and the Fundamental Research Funds for the Central Universities. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Competing Interests: The authors have declared that no competing interests exist. Multi-voxel pattern analysis (MVPA) methods have been widely used in fMRI studies to characterize the relationship between fMRI responses and cognitive functions. In a typical MVPA framework, a multivariate classifier is trained on multi-voxel patterns of brain activities with known labels, and the trained classifier is then used to classify untrained data (for a review, see [1,2,3]). The MVPA methods have been proved successful in differentiating stimulus categories or task states in both human [4,5,6,7,8,9] and monkey fMRI studies [10]. Recently, interests have been dedicated to reliably localizing such discriminative information in the brain [3]. Here we proposed a new approach based on clusters rather than voxels for this purpose. One of the core objectives in functional neuroimaging is to link cortical regions to cognitive functions. Accordingly, a univariate statistical parametric mapping approach has been proposed [11,12]. However, the univariate approach is unable to extract information embedded in multi-voxel patterns because each voxel is treated independently. To address this issue, multivariate feature selection algorithms from machine learning have been used in analyzing multi-voxel patterns of brain activation [13,14,15, 16,17,18]. These algorithms mainly aim to select a minimum set of voxels necessary for constructing a classifier with the best predictive accuracy [19]. However, because extensive spatiotemporal correlations in fMRI responses among neighboring voxels lead to high redundancy of features, only a small set of voxels with similar response profiles are selected. This leads to two problems that limit the application of voxel-based MVPA methods in functional brain mapping. First, the selected voxels are usually distributed rather than clustered, some of which may be present outside the brain (e.g. [15], but see [16,18]). Second, small variations of data may cause completely different sets of voxels being selected [17]. To tackle these problems, we used local homogeneous clusters [20,21], not individual voxels, as basic units for the brain mapping in the MVPA. To this end, we proposed a new approach where both brain activation patterns within local homogeneous clusters and the interaction among these clusters were examined. We termed this approach as mapping informative clusters (MIC), in contrast to traditional voxel-based approaches of mapping informative voxels (MIV). Results from both simulated and real fMRI data showed that our MIC method outperformed the MIV method in localizing informative regions while the predictive accuracy was largely preserved. In the MIC, a hierarchical framework was used to identify the most informative clusters by examining both local and global patterns of fMRI data (Figure 1). First, through a multi-voxel classifier, the multi-voxel pattern within a homogeneous cluster Figure 1. The framework of the mapping informative clusters (MIC) approach. The whole brain was partitioned into homogeneous clusters, where the within-cluster voxel patterns from different clusters were transformed into a cluster pattern consisted of one single value for each cluster. Informative clusters were selected according to their discriminative weights derived from the cluster pattern. doi:10.1371/journal.pone.0015065.g001 was summarized to a single value that represented conditionrelated information carried by the cluster. The output values of all clusters were then taken as a multi-cluster pattern to construct a second-level multi-cluster classifier so as to yield cluster weights for cluster ranking. Finally, informative clusters were selected based on the ranking of all clusters. A cross-validation procedure was used to evaluate the performance of different mapping methods. That is, in each fold of the cross-validation, the original data were split into training and test data sets. The training data set was used for both selecting informative voxels (MIV) or clusters (MIC) and constructing a linear SVM classifier for classifying experimental conditions from the selected clusters. The test data set was then used to evaluate the predictive accuracy of the classifier constructed from the training data set. Partition of homogeneous clusters Because voxels within a homogeneous cluster by definition show similar response profiles, we used an iterative algorithm of competitive region growing [21,22] to partition the whole brain into homogeneous clusters. The similarity between (...truncated)