SARDU-Net

SARDU-Net selects informative subsets of quantitative MRI (qMRI) measurements using a data-driven, model-free deep learning framework to design economical, high-information qMRI protocols.

Key Features:

• Data-driven, model-free approach: Operates without predefined biophysical models or tissue-parameter optimization to enable protocol selection across anatomical contexts.
• Deep learning architecture: Employs two deep neural networks trained jointly in an end-to-end manner.
• Selector network: Identifies the most informative fixed-size subset of qMRI measurements from lengthy pilot scans.
• Predictor network: Reconstructs or estimates fully-sampled qMRI signals from the selected subset via direct upsampling.
• Fixed-size sub-protocol selection: Selects reproducible, fixed-size sub-protocols from pilot scans to maximize information content.
• Efficient selection from exhaustive pilot scans: Enables selection of informative measurements without exhaustive manual search.
• Compatibility with downstream mapping: Selected sub-protocols have been shown to produce robust T1-weighted spherical mean diffusion tensor (T1-SMDT) and hybrid multi-dimensional MRI (HM-MRI) maps with high goodness-of-fit.

Scientific Applications:

• Brain imaging (DRI): Applied to diffusion-relaxation imaging (DRI) in the brain using 32 diffusion-/T1-weightings to generate T1-SMDT maps.
• Prostate imaging (DRI): Applied to diffusion-relaxation imaging in the prostate using 16 diffusion-/T2-weightings for HM-MRI modeling.
• Validation on 3T systems: Tested on scans from healthy volunteers acquired on two separate 3T Philips MRI systems.

Methodology:

Jointly train a selector and a predictor deep neural network end-to-end on pilot qMRI scans; the selector chooses fixed-size measurement subsets and the predictor reconstructs fully-sampled signals from those subsets via direct upsampling.