TRD 3: Removing the biological limits for in vivo human brain MRI

While important advances in hardware and software for image acquisition have advanced neuroMRI significantly in recent years (and will continue to do so; see TRDs 1, 2), their success has opened the door to a new set of challenges – many of the most fundamental limitations are now set by human biology, including vascular limits to fMRI spatial resolution, nuisance physiological modulation, and SAR and Peripheral Nerve Stimulation (PNS) concerns with advanced RF and gradient waveforms. Thus, in TRD3, we propose a program of bioengineering development to bridge the macro, meso and micro scales of brain architecture, by improving the spatio-temporal resolution of fMRI down to its biological limits. Overall, our tools will advance the study of brain circuits in living humans at mesoscopic scales but extending over hemispheric-sized regions. We propose improving the MR acquisition on multiple levels. Since MR encoding is first and foremost limited by gradient performance, we will develop gradient coil design and sequence approaches that directly incorporate Peripheral Nerve Stimulation (PNS) barriers into their optimization. This is needed to advance beyond the performance of modern 80mT/m (and higher) gradients where fMRI is biologically limited (by PNS) rather than the engineering issues of previous gradient generations (efficiency, power or heat removal). Secondly, we will advance acquisition strategy to encode biological nuisance modulation of k-space data from respiration and motion, modeling these effects on the individual’s data, then mitigating it in the reconstruction process using a machine learning (ML) approach. Finally, we will address the biological limit imposed on fMRI spatial resolution by the hemodynamic response which degrades high-resolution functional MRI maps. We will use an ML approach to remove the “filter” imposed by the vasculature on the activation pattern. We will train a Convolutional Neural Net (CNN) to learn the relationship between observed high-resolution fMRI patterns and the ground-truth activation pattern (known from retinotopic activation studies in primary visual cortex) incorporating a priori measured micro-vasculature patterns seen in high-resolution arteriograms and venograms. After training and validating the CNN using the known pattern projected on the retina, we will test its generalization to “un-do” the neurovascular “filter” in V1 retinotopy, other V1 features such as Ocular Dominance Column (ODC) maps, higher visual areas such as V2, and finally in non-visual areas such as somatosensory and auditory cortex.