Skip to Main Content
Data-driven cluster analysis is potentially suitable to search for, and discriminate between, distinct response signals in blood oxygenation level-dependent functional magnetic resonance imaging (BOLD fMRI), which appear during cerebrovascular disease. In contrast to model-driven methods, which test for a particular BOLD signal whose shape must be given beforehand, data-driven methods generate a set of BOLD signals directly from the fMRI data by clustering voxels into groups with correlated time signals. Here, we address the problem of selecting only the clusters that represent genuine responses to the experimental stimulus by modeling the correlation structure of the clustered data using a Bayesian hierarchical model. The model is empirically justified by demonstrating the hierarchical organization of the voxel correlations after cluster analysis. BOLD signal discrimination is demonstrated using: 1) simulations that contain multiple pathological BOLD response signals; and 2) fMRI data acquired during an event-related motor task. These demonstrations are compared with results from a model-driven method based on the general linear model. Our simulations show that the data-driven method can discriminate between the BOLD response signals, while the model-driven method only finds one signal. For fMRI, the data-driven method distinguishes between the BOLD signals appearing in the sensorimotor cortex and those in basal ganglia and putamen, while the model-driven method combines these signals into one activation map. We conclude that the proposed data-driven method provides an objective framework to identify and discriminate between distinct BOLD response signals.