Machine Learning-Enabled High-Resolution Dynamic Deuterium MR Spectroscopic Imaging.

Area of Science:

• Biophysics
• Medical Imaging
• Machine Learning

Background:

• Deuterium magnetic resonance spectroscopic imaging (DMRSI) shows promise for noninvasive imaging of brain metabolism and tumors.
• Low sensitivity currently limits DMRSI's research and clinical utility.

Purpose of the Study:

• To develop a novel machine learning-based method to enhance DMRSI sensitivity and enable high-resolution imaging.
• To address the limitations of low signal-to-noise ratio in dynamic DMRSI.

Main Methods:

• A hybrid approach combining physics-based subspace modeling and deep neural networks for signal denoising.
• Training deep neural networks to learn low-dimensional spectral and temporal manifolds of DMRSI data.
• Integrating learned subspace and manifold structures via regularization to remove noise.

Main Results:

• The proposed method effectively denoises DMRSI data, enabling high-resolution imaging.
• Demonstrated denoising efficacy through theoretical analysis, simulations, and in vivo experiments.
• Successfully captured the Warburg effect and tumor heterogeneity in preclinical models.

Conclusions:

• The novel machine learning method significantly enhances DMRSI sensitivity and imaging capabilities.
• This approach holds translational potential for basic research and clinical applications in oncology and neuroscience.
• The framework may be applicable for denoising other spatiospectral data types.