A deep network for tissue microstructure estimation using modified LSTM units.

Area of Science:

• Neuroimaging
• Biomedical Engineering
• Computational Neuroscience

Background:

• Diffusion magnetic resonance imaging (dMRI) is crucial for noninvasively assessing tissue microstructure.
• Accurate microstructure estimation is challenging with limited diffusion gradients due to complex signal models.
• Deep learning, particularly optimization-based learning, shows promise for improving dMRI microstructure estimation.

Purpose of the Study:

• To propose a novel deep network for tissue microstructure estimation in dMRI using optimization-based learning.
• To incorporate historical information into the deep network design for enhanced iterative optimization.
• To improve the accuracy of microstructure estimation with a reduced number of diffusion gradients.

Main Methods:

• Developed a two-stage deep network based on unfolding iterative sparse reconstruction processes.
• Incorporated historical information within the network structure, analogous to modified Long Short-Term Memory (LSTM) units.
• Jointly learned weights by minimizing mean squared error for microstructure estimation using spatial-angular sparse representation.

Main Results:

• The proposed optimization-based learning network demonstrated superior performance in microstructure estimation compared to existing methods.
• Effective estimation was achieved even with a reduced number of diffusion gradients.
• Validated on neurite orientation dispersion and density imaging (NODDI), spherical mean technique (SMT), and ensemble average propagator (EAP) models.

Conclusions:

• The novel deep network effectively leverages historical information for improved dMRI microstructure estimation.
• This approach offers a significant advancement for accurate tissue microstructure analysis in dMRI, especially in data-limited scenarios.
• The method shows broad applicability across various dMRI signal models.