Deep Learning-Based Diffusion MRI Tractography: Integrating Spatial and Anatomical Information

Motivation: Tractography faces persistent issues with false positives and negatives. Studies suggest that adding novel information during reconstruction may improve sensitivity-specificity. Goal(s): Our objective is to accurately predict propagation directions by integrating image-domain neighborhood information and contextual information along streamlines. Approach: A network combining convolutional layer and several Transformer-decoders is proposed to integrate novel information. Results: On the in-vivo dataset, our method achieves an average improvement of 5% in white matter coverage compared to existing methods, while maintaining a minimal increase of only 1% in overreach. On ISMRM2015 Challenge dataset, our method reconstructs 24 out of 25 bundles with 66% valid streamlines. Impact: The proposed method successfully generates anatomically plausible streamlines across both synthetic and in-vivo human brain datasets. These promising results suggest that exploring additional novel information could further improve the anatomical reliability of white matter mapping.