Current Challenges and Future Directions in Diffusion MRI: From Model- to Data- Driven Analysis

Diffusion weighted MRI is a prominent non-invasive modality to probe in vivo tissue micro- and macro-structure and has been widely applied throughout neuro- and body imaging. The promise of micro-scale analyses has been in the creation of virtual biopsies that provide information in place of physical histology, while tractography and its related methods offer maps of the neuronal wiring through virtual dissection. While both approaches have had strong successes at the group level, specificity and sensitivity at the individual dataset/single subject level have been more elusive. Herein, we reflect on current challenges and potential future directions in the context of a futurist piece. As such, we go beyond the reasonably well-established science to offer hypotheses/postulates/challenges to encourage discussion and exploration. We postulate that there are transformative opportunities available if we complement our perspective of diffusion MRI as a signal that is explained by a tractable biophysical model with one in which data driven machine learning can inform us about detection, localization, and assessment of both normal and abnormal brain tissue in both local (voxels) and global connectivity. Towards this end, this manuscript describes challenges associated with achieving virtual biopsy (i.e., microstructural modeling) and virtual dissection (i.e., fiber tractography) and suggests opportunities to use data-driven techniques to improve modeling geometry, to learn features of the signal that may prove useful as biomarkers, and to harmonize signal, techniques, and datasets to improve tissue characterization.