Approximate diffusion tractography from FLAIR MRI and anatomical context using recurrent neural networks

Diffusion MRI (dMRI) tractography methods provide a valuable method for in-vivo estimation of whole-brain white matter pathways that are commonly assumed to rely on microstructure models derived from dMRI. However, recent pioneering works have demonstrated that the accuracy of white-matter measurements computed from T1-weighted (T1w) MRI tractography is on a similar level to scan-rescan variability in dMRI tractography. This revelation raises new questions about understanding tractography: Is it primarily a dMRI microstructural phenomenon, and how different can it be when estimated from other imaging modalities? In this study, we propose a framework to approximate tractography from fluid-attenuated inversion recovery (FLAIR) MRI and examine its performance compared to tractography based on diffusion and T1w MRI. We adapt the teacher-student recurrent neural network (RNN) model from existing work on T1w tractography. Additionally, we use brain segmentation maps as the anatomical context. We conduct white matter bundle analysis and compare various metrics with those from T1w tractography and the traditional dMRI tractography. FLAIR tractography achieved significant different performance compared to T1 tractography evaluated by Dice similarity coefficient (p=0.004) and bundle adjacency streamlines distance (p=0.012). An average absolute difference of 23% was observed in eight bundle shape measurements between FLAIR tractography and traditional dMRI tractography. Both qualitative and quantitative results suggest that tractography based on FLAIR MRI is feasible and underscore the need for comprehensive research to understand tractography in the broader context of multi-modality brain MRI.