Super-Resolved q-Space deep learning with uncertainty quantification.

Area of Science:

• Neuroimaging
• Medical Physics
• Machine Learning

Background:

• Diffusion magnetic resonance imaging (dMRI) noninvasively measures brain tissue microstructure.
• Existing q-Space deep learning (q-DL) methods estimate microstructure from reduced dMRI data.
• Limitations in dMRI include reduced gradient numbers and low spatial resolution, impacting microstructure estimation quality.

Purpose of the Study:

• To extend q-DL for super-resolved tissue microstructure estimation using low-resolution dMRI signals.
• To introduce super-resolved q-DL (SR-q-DL) for mapping low-resolution signals to high-resolution microstructure.
• To develop probabilistic SR-q-DL for quantifying estimation uncertainty and improving accuracy.

Main Methods:

• Developed SR-q-DL using deep networks for patch-based mapping from low-resolution to high-resolution microstructure.
• Integrated diffusion signal sparsity into network design with sparse representation and mapping components.
• Proposed probabilistic SR-q-DL using a deep ensemble strategy for uncertainty quantification and improved accuracy.

Main Results:

• SR-q-DL demonstrated superior estimation accuracy compared to competing methods on two independent dMRI datasets.
• Probabilistic SR-q-DL successfully quantified estimation uncertainty.
• Uncertainty measures correlated with estimation errors, indicating potential for brain studies.

Conclusions:

• SR-q-DL effectively enhances dMRI-based tissue microstructure estimation by increasing spatial resolution.
• Probabilistic SR-q-DL provides reliable uncertainty quantification alongside accurate microstructure mapping.
• The proposed methods offer significant advancements for quantitative analysis in brain imaging research.