An improved deep network for tissue microstructure estimation with uncertainty quantification.

Area of Science:

• Neuroimaging
• Biophysics
• Machine Learning

Background:

• Diffusion magnetic resonance imaging (dMRI) is vital for studying brain tissue microstructure.
• Deep learning methods have advanced dMRI analysis, especially with reduced gradient acquisitions.
• Existing deep learning approaches face challenges with large network parameters and lack uncertainty quantification.

Purpose of the Study:

• To develop an improved deep learning network for accurate tissue microstructure estimation from sparse dMRI data.
• To address limitations of existing methods, including large parameter counts and lack of uncertainty estimation.
• To introduce a novel uncertainty quantification strategy for deep learning-based dMRI analysis.

Main Methods:

• Explored sparse representation of diffusion signals using a separable spatial-angular dictionary.
• Designed and unfolded an iterative deep network based on sparse code updates.
• Implemented a residual bootstrap strategy, specifically Lasso bootstrap, for quantifying estimation uncertainty.

Main Results:

• The proposed deep network demonstrated favorable estimation accuracy compared to state-of-the-art methods on brain dMRI data.
• The uncertainty measures generated by the method correlated well with estimation errors.
• The approach provided reasonable confidence intervals, indicating potential for clinical brain studies.

Conclusions:

• The developed deep learning network effectively estimates tissue microstructure from undersampled dMRI data.
• The novel uncertainty quantification method provides reliable confidence intervals for dMRI-derived microstructure estimates.
• This work offers a promising tool for advancing quantitative analysis in neuroimaging research.