tDKI-Net: A Joint q-t Space Learning Network for Diffusion-Time-Dependent Kurtosis Imaging.

Area of Science:

• Magnetic Resonance Imaging
• Biophysics
• Computational Biology

Background:

• Time-dependent diffusion magnetic resonance imaging (TDDMRI) enables non-invasive tissue microstructure characterization.
• Current TDDMRI models necessitate densely sampled q-t space data, leading to lengthy acquisition protocols.
• Accelerating TDDMRI acquisition is crucial for clinical applicability and broader research.

Purpose of the Study:

• To develop and validate tDKI-Net, a novel deep learning model for estimating diffusion-time dependent kurtosis and transmembrane exchange using undersampled q-t space data.
• To significantly reduce TDDMRI scan times without sacrificing estimation accuracy.
• To improve the efficiency of microstructural characterization in biological tissues.

Main Methods:

• Proposed tDKI-Net, a joint q-t space model integrating q-Encoders and a t-Encoder based on the extragradient mechanism.
• Employed a three-stage training strategy: physics-informed self-supervised pretraining, DKI warm-up, and joint training.
• Utilized downsampled q-t space data for model training and validation, comparing performance against existing methods.

Main Results:

• tDKI-Net effectively estimated diffusion-time dependent kurtosis and transmembrane exchange time from undersampled data.
• The proposed three-stage training strategy outperformed training from scratch, reducing NRMSE for by up to 1.4%.
• Achieved lower NRMSEs for , , and (2.50%, 3.04%, 10.86%) with one-subject training compared to prior three-subject training studies.
• Reduced scan time by 10.5-fold through joint q-t space data downsampling without compromising accuracy.

Conclusions:

• tDKI-Net offers a significant acceleration of TDDMRI acquisition protocols.
• The model accurately estimates key microstructural parameters, including transmembrane exchange time.
• This approach holds promise for more efficient and accessible diffusion MRI-based tissue characterization.