Jointly estimating parametric maps of multiple diffusion models from undersampled q-space data: A comparison of three...

Area of Science:

• Neuroimaging
• Medical Physics
• Computational Neuroscience

Background:

• Advanced diffusion MRI techniques offer valuable insights into brain microstructure.
• Clinical adoption is limited by long scan times and the need for multiple models.
• Current methods extract limited features per technique, necessitating multiple complex model fittings.

Purpose of the Study:

• To compare deep learning (DL) approaches for jointly estimating parametric maps of multiple diffusion models.
• To assess DL performance in extracting microstructural features from undersampled q-space data.
• To evaluate the efficiency and accuracy of DL methods in advanced diffusion MRI.

Main Methods:

• Implementation of three DL approaches: 1D-qDL, 2D-CNN, and MESC-SD.
• Joint estimation of parametric maps for diffusion tensor imaging (DTI), diffusion kurtosis imaging (DKI), neurite orientation dispersion and density imaging (NODDI), and spherical mean technique (SMT).
• Utilizing highly undersampled q-space data for model fitting.

Main Results:

• The MESC-SD network architecture demonstrated the highest accuracy, with normalized RMSE below 10% in most brain regions.
• Accuracy is influenced by q-space undersampling, network architecture, and specific brain regions/parameters.
• DL methods enable simultaneous estimation of multiple diffusion maps.

Conclusions:

• DL methods provide efficient tools for simultaneous estimation of various diffusion maps from undersampled data.
• Significant reductions in scan time (6-fold) and processing time (25-fold) were achieved.
• These DL approaches offer a practical solution for obtaining advanced diffusion parametric maps with reasonable accuracy.