Deep Learning-based Fast Calculation of Diffusion Tensor Distribution Parameters

Motivation: Current DTD methods rely on Monte Carlo inversion, which is computationally expensive and time-consuming, and treat each voxel independently, neglecting spatial relationships between neighboring voxels. Goal(s): To accelerate the reconstruction of DTD parameter maps using deep learning, while maintaining accuracy. Approach: The model estimates voxel-specific diffusion tensor parameters using spatial correlations in diffusion MRI data. It combines a transformer's self-attention mechanism for encoding spatial dependencies and sparse dictionary learning for decoding. Results: DTDMapper is a deep learning model that accelerates DTD parameter map reconstruction by over 100 times without sacrificing accuracy, showing high performance across various pathologies for broader disease applicability. Impact: This study introduces deep learning for DTD parameter computation, overcoming computational complexity and spatial continuity issues of Monte Carlo methods, with potential for clinical translation across various diseases.