DIMOND: DIffusion Model OptimizatioN with Deep Learning.

Area of Science:

• Neuroimaging
• Biophysics
• Computational Neuroscience

Background:

• Diffusion magnetic resonance imaging (dMRI) is crucial for non-invasive in vivo human brain mapping.
• Accurate estimation of dMRI model parameters is computationally intensive and sensitive to noise.
• Existing supervised deep learning methods require extensive training data and may lack generalizability.

Purpose of the Study:

• To introduce DIMOND, a novel physics-informed, self-supervised deep learning framework for dMRI parameter estimation.
• To address the computational cost and generalizability limitations of current dMRI analysis techniques.
• To improve the efficiency and accuracy of mapping brain tissue microstructure and structural connectivity.

Main Methods:

• DIMOND utilizes a neural network to map dMRI image data to diffusion model parameters.
• The network is optimized by minimizing the difference between acquired and synthetically generated dMRI data.
• Physics-informed and self-supervised learning principles guide the optimization process.

Main Results:

• DIMOND achieves accurate diffusion tensor imaging (DTI) results.
• The framework demonstrates generalizability across different subjects and datasets.
• DIMOND outperforms conventional methods for complex models like kurtosis and Neurite Orientation Dispersion and Density Imaging (NODDI).
• Transfer learning with DIMOND significantly reduces NODDI model fitting time from hours to seconds.

Conclusions:

• DIMOND offers a highly effective and efficient solution for dMRI parameter estimation.
• The self-supervised nature of DIMOND enhances its practical feasibility for clinical and neuroscientific applications.
• DIMOND facilitates broader adoption of microstructure and connectivity mapping in research and healthcare settings.