Yoon, Jongyeon; Rao, Mingxing; McMaster, Elyssa M.; Cho, Chloe; Newlin, Nancy R.; Schilling, Kurt G.; Landman, Bennett A.; Moyer, Daniel. “Transformer-based T1-tractography.” Proceedings – International Symposium on Biomedical Imaging (2025). https://doi.org/10.1109/ISBI60581.2025.10981144.
Diffusion MRI (dMRI) is currently the best non-invasive way to map white matter pathways in the human brain—these are the “wiring” that connects different brain areas. However, dMRI isn’t always available, while another type of brain scan, called T1-weighted (T1w) MRI, is much more commonly used.
Recently, researchers have started using deep learning to try and recreate these brain connection maps (called streamlines) from T1w images instead of dMRI. One of the best current methods for doing this is called CoRNN, but it uses a type of neural network (a “recurrent” structure) that limits its accuracy.
In this study, we improved CoRNN by replacing those recurrent parts with a more modern architecture known as Transformers. We also updated how the method represents and predicts the directions of brain fibers. These upgrades led to better results, showing that our new approach more closely matches the gold standard dMRI maps in healthy adults.
Fig. 1.
Framework for training proposed method in teacher-student framework. The GRUs were replaced with a transformer model to better capture long-range dependencies, and the single linear layer predictor for spherical coordinates was replaced with an MLP that predicts unit vectors in cartesian space. SAMP refers to sampling the grid using trilinear interpolation.
