Spatio-temporal fuzzy clustering of functional magnetic resonance imaging data

Magnetic resonance imaging (MRI) is a preferred imaging modality due to its high resolution images of in vivo tissue. Functional MRI (fMRI) infers organ function using blood flow intensities. However, multiple response models for hemodynamics and, more specifically, neural activation, contend for widespread adoption. Development of models, imaging techniques and various types of noise compound problems in analysis and motivate the use of exploratory data analysis to elicit intrinsic data structure. This work demonstrates the utility and efficacy of a novel exploratory data analysis technique derived from a robust, unsupervised learning method, fuzzy C-means (FCM). The algorithm, designated FCM with feature partitions (FCMP), integrates feature relationships in the clustering process. One feature relation not widely exploited in fMRI analysis is the high probability that temporally similar time courses are also spatially proximal. FCMP has exploited this relation to generate both novel and robust data inferences. Both synthetic and in vivo fMRI data are examined. FCMP is compared to benchmarks from industry and academia, including FCM, cluster merging, CHAMELEON and EvIdentRTM. Ten distinct experiments examine aspects of FCMP with respect to fMRI analysis, in particular, means to integrate distinct feature subsets and feature relationships, sample membership in regions of interest, use of validation indices for fMRI, and data-driven global thresholding. Efficacy of FCMP for fMRI analysis is shown in terms of noise reduction, statistical specificity, and discovery of novel spatial relations between time courses in regions of interest.