Presentation

Search Abstracts | Symposia | Slide Sessions | Poster Sessions | Lightning Talks

Using Machine learning to localize Area Spt without an Spt localizer.

Poster B89 in Poster Session B, Tuesday, October 24, 3:30 - 5:15 pm CEST, Espace Vieux-Port

Bradley Buchsbaum¹; ¹Rotman Research Institute

In the field of cognitive neuroscience of language, reliable identification of functional regions of interest is crucial for obtaining comparable findings across research studies, laboratories, and neuroimaging parameters and pipelines. In the early 2000s, numerous groups used fMRI to localize functionally specific regions of the brain thought to underlie specific "cognitive processes". Through a series of fMRI investigations focused on mapping regions jointly involved in speech perception, speech production, and verbal short-term memory, Buchsbaum et al. (2001) and Hickok et al. (2003) targeted a functional region at the posterior end of the left Sylvian fissure (Area Spt - Sylvian-parietal-temporal). This area was hypothesized to act as a bridge or interface between the perception and production of speech. Over a decade later, Glasser et al. (2016) used large-scale multimodal mapping of neuroimaging data from the Human Connectome Project to partition the entire cerebral cortex into 360 regions. One of these regions, the perisylvian language area (PSL), was anatomically co-located with the functionally-defined area Spt. Here, we present results from 15 research subjects scanned with fMRI, who underwent a 10-minute resting state scan, viewed 20 minutes of movies, and performed the classic Spt localizer (auditory-verbal short-term memory). We then trained machine learning models with spatial priors to identify the boundaries of Spt using functional connectivity patterns from the movie and resting state data. Preliminary results indicate that Spt can be localized using connectivity data derived from as little as 5 minutes of resting state data, with prediction accuracy increasing with more training data. We also examined the extent to which a group-defined PSL ROI overlapped with funcitonally-defined Spt. The results suggest rough agreement in the group-averaged data, but with significant variability across subjects. These data open up the possibility of localizing Spt in individual subjects without an explicit "Spt localizer"--and without falling back on an approximate group parcellation--thereby enabling analysis of Spt responses in a variety of historical, openly shared, or resting state-only fMRI datasets.

Topic Areas: Computational Approaches, Multisensory or Sensorimotor Integration