Individualised mapping of the language system using ultra high-field fMRI

2408 

Abstract Submission 

Nicole Eichert¹, Donna Gift Cabalo², Yezhou Wang², Shahin Tavakol³, Jordan DeKraker⁴, Alexander Weil⁵, Raúl Rodriguez-Cruces⁶, Boris Bernhardt⁶

¹University of Oxford, Oxford, Oxfordshire, ²McGill University, Montreal, Quebec, ³McGill University, Montreal Neurological Institute and Hospital, Montreal, Quebec, ⁴McGill University, Montreal, Canada, ⁵Centre Hospitalier Universitaire Sainte-Justine, Montreal, Canada, ⁶Montreal Neurological Institute and Hospital, Montreal, Quebec

Nicole Eichert  
University of Oxford
Oxford, Oxfordshire

Donna Gift Cabalo  
McGill University
Montreal, Quebec

Yezhou Wang  
McGill University
Montreal, Quebec

Shahin Tavakol  
McGill University, Montreal Neurological Institute and Hospital
Montreal, Quebec

Jordan DeKraker  
McGill University
Montreal, Canada

Alexander Weil  
Centre Hospitalier Universitaire Sainte-Justine
Montreal, Canada

Raúl Rodriguez-Cruces  
Montreal Neurological Institute and Hospital
Montreal, Quebec

Boris Bernhardt  
Montreal Neurological Institute and Hospital
Montreal, Quebec

All humans share the fundamental neurobiological infrastructure that enables us to process language. Nevertheless, the brain's language system exhibits considerable inter-individual variability in both anatomy and function [1]. Mapping the language system, therefore, requires a nuanced approach that captures neuroanatomical features with high reliability to enable individualised analyses. Here, we performed high-precision individualised mapping of the language system using multimodal ultra high-field MRI. We leveraged a new MRI resource with unprecedented coverage: The openly-available 7T Precision NeuroImaging (PNI) dataset from the Montreal Neurological Institute features an extensive battery of repeated structural, quantitative and functional scans. Each participant underwent 4 sessions, one of which included an fMRI language task. This comprehensive dataset allows us to localise the brain's language system and to test associations with intrinsic neuroanatomical features.

The imaging protocol was implemented on a 7T Siemens Terra scanner. Pre-processed data from six participants (4F, median age: 25 years [22-32]) were accessed from the PNI dataset, specifically task BOLD fMRI data (1.9mm isotropic, TR=1690ms, multiband factor 3, echo spacing=0.53ms, multi-echo) and resting-state fMRI data (6 mins) from the same session. From other scanning sessions we accessed apparent diffusion coefficient (ADC), magnetization transfer saturation (MTSAT) and MP2RAGE-based quantitative T1 relaxometry (qT1). All metrics were sampled to individual surfaces reconstructed using FastSurfer and manually corrected. Pre-processing was performed using Micapipe [2]. The language task design and stimuli, matched for linguistic features, were adapted from a previous study [3]. Briefly, in blocks of 30s (10 trials of 3 seconds), participants were required to make a phonological, semantic, or visual judgement about a pair of words (Figure 1A). Task data were analysed using Python's nilearn to obtain a phonological-versus-semantic contrast. Group-level activation maps were based on t-test from individual z-maps and based on thresholded activation count maps. Group-level task peaks were identified on the surface and individual task peaks were located within a circular ROI of 2cm. To demonstrate specificity, random (i.e., displaced) peaks were selected within the same ROIs. Functional connectivity (FC) was determined using Pearson's correlation from rs-fMRI data. Associations of individual z-maps to microstructural markers were tested using a linear mixed effects model as implemented in Python's statsmodels.

Behavioural responses indicated that all participants showed high performance throughout the task and comparable difficulty of phonological and semantic conditions (Figure 1B). Group-level activation and count maps demonstrate that expected language hubs in the brain were engaged during phonological and semantic processing with overall left-ward lateralization (Figure 1C). Group-level FC analyses demonstrated that phonological and semantic processing are relying on two distinct networks, which can be reliably extracted on the individual subject level (Figure 2A). A small displacement of the individual peaks results in a degraded connectivity matrix indicating high spatial specificity of the language task localizer. Spatial correlation to microstructural markers showed small but consistent associations across subjects, with semantic processing co-localising with cortex with lower ADC, lower MTSAT, and higher qT1 (Figure 2B).

We successfully mapped the language system using a robust localizer task at 7T maintaining individual subject differences. The task scan is part of a densely sampled and multimodal MRI dataset, which will be openly available for the community. In our exemplary analysis on structure-function relations within the cortex, we showed that phonological and semantic processing are supported by distinct neuroanatomical make-up of the cortex.

Language Comprehension and Semantics ²

Anatomical MRI

Multi-Modal Imaging ¹

Imaging Methods Other

Experimental Design

FUNCTIONAL MRI

HIGH FIELD MR

Language

Open Data