MINeR: Direction-modulated implicit neural representation enables ultrafast multi-shell diffusion MRI

Area of Science:

• Medical Imaging
• Neuroscience
• Biophysics

Background:

• Diffusion MRI (dMRI) maps tissue microstructure via water diffusivity.
• Advanced dMRI models offer cellular sensitivity but require long acquisition times due to dense q-space data.
• Current deep learning methods for dMRI parameter estimation are limited by fixed protocols, model assumptions, and poor generalization.

Purpose of the Study:

• To introduce MINeR, a novel unsupervised framework for reconstructing dense q-space dMRI data from undersampled acquisitions.
• To enable accurate microstructural parameter estimation for diverse diffusion models using sparse data.
• To improve the clinical applicability of dMRI by reducing acquisition time.

Main Methods:

• Developed MINeR, an unsupervised, subject-specific framework utilizing direction-modulated implicit neural representation.
• Employed MINeR to flexibly sample diffusion signals across q-space from highly undersampled data.
• Evaluated MINeR's ability to reconstruct dense q-space data and estimate microstructural parameters.

Main Results:

• MINeR reconstructs high-quality dMRI signals from as few as 6 directions, significantly reducing scan time.
• The framework maintains high fidelity in microstructural parameter estimation, especially for advanced multi-shell models.
• MINeR demonstrates robust generalization capabilities, performing well on tumor data.

Conclusions:

• MINeR provides a practical approach for microstructural modeling from sparsely sampled dMRI data.
• The method significantly reduces acquisition time while preserving robust parameter estimation.
• This framework enhances the clinical utility of diffusion MRI by overcoming data acquisition limitations.