Application of functional magnetic resonance imaging to the study of language

Javier Gonzalez Castillo, Purdue University


Functional Magnetic Resonance Imaging (fMRI) is a powerful neuroimaging technique that allows in-vivo imaging of brain function using local changes in blood oxygenation levels (BOLD) as an intrinsic metabolic marker of neural activity. This dissertation presents a series of fMRI studies designed to advance our understanding of the neural correlates of language processing, focusing on three important aspects: (1) neural correlates of pre-lexical analysis of normal and spectrally degraded speech—e.g., cochlear implant speech; (2) neural correlates of conceptual representations of action verbs; and (3) design of robust, highly reliable fMRI experimental protocols suitable for longitudinal and clinical studies of speech function.^ The first study focused on understanding the neural correlates of adaptation to acoustic simulations of spectrally degraded speech that mimics what cochlear implant (CI) users perceive through their implants. Six normal hearing subjects underwent fifteen sessions of training with degraded speech produced using an acoustic CI simulator [2]. Two fMRI imaging sessions were conducted, one prior and one subsequent to training. During each imaging session subjects were presented with both normal and degraded speech utterances while performing a non-linguistic discrimination task. It was found that training in degraded speech perception was associated with increased recruitment of the bilateral posterior cingulate cortex, the right inferior frontal gyrus (BA 46), and the left inferior frontal junction, as well as reduced activity within a region in the posterior Sylvian fissure at the parietal-temporal boundary.^ In respect to the neural correlates of conceptual representations, an fMRI study was conducted to investigate a series of hypothesis about the neural representations of five semantic categories (e.g., action, motion, contact, change of state, tool use) derived from the Simulation Framework of L.W. Barsalou [3]. Based on the Simulation Framework, we hypothesized that the bodily action component depends on the primary motor and premotor cortices, that the visual motion component depends on the posterolateral temporal cortex, that the contact component depends on the intraparietal sulcus and inferior parietal lobule, that the change of state component depends on the ventral temporal cortex, and that the tool component depends on a distributed network of temporal, parietal, and frontal regions. Virtually all of these predictions were confirmed. As a group, these findings support the Simulation Framework and extend our understanding of the neuroanatomical distribution of different aspects of verb meaning.^ As for the reproducibility of fMRI activations associated with language comprehension, we conducted a third fMRI study to evaluate the test-retest reproducibility of fMRI activations associated with auditory sentence comprehension under the absence of experimentally controlled manipulations. A bilateral neural network distributed across lateral temporal cortex, dorsolateral frontal cortex, medial supplementary motor cortex (medial BA6), visual cortex, adjacent cerebellar cortex, and basal ganglia was consistently activated at all three levels. This network is believed to be involved in speech comprehension—including auditory, lexical and semantic processing—, short term memory encoding, and attention control. Reproducibility of three different aspects of these activations—namely binary activation maps, magnitude, and spatial distribution of local maxima—was evaluated in order to assess if a single pre-manipulation session provides a good characterization of the typical activations associated with the task when all experimentally controllable variables are kept constant. At the group level, activations proved highly reliable with 83.95% of the imaged volume being consistently active/inactive across all five sessions (average ratio of volume overlap = 0.79 ± 0.01; average Pearson's correlation = 0.93 ± 0.02). At the single-subject level, consistency of activations is considerably lower with an average of 58.65% of the volume being consistently classified for a given subject (average ratio of volume overlap = 0.68 ± 0.09; average Pearson's correlation = 0.86 ± 0.07). In addition, two different types of intraclass correlation coefficients were implemented to identify areas with high ratios of between-subject and between-session variance. Finally a newly defined metric, the ninety five percent probability radius (r95), that measures reproducibility of spatial localization, was used to identify the most spatially reproducible peak locations across group and single-subject results and provide an estimate of the random shift that can be expected even in the absence of any experimental manipulation. (Abstract shortened by UMI.) ^




Thomas M. Talavage, Purdue University.

Subject Area

Biology, Neuroscience|Engineering, Biomedical

Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server