A deep learning approach to estimation of subject-level bias and variance in high angular resolution diffusion...

Area of Science:

• Neuroimaging
• Medical Physics
• Computational Neuroscience

Background:

• Diffusion MRI (dMRI) quality assessment is crucial for reliable neuroimaging analysis.
• Estimating bias and variance of generalized fractional anisotropy (GFA) improves inference and replicability.
• Existing methods for GFA bias and variance estimation are computationally complex.

Purpose of the Study:

• To develop fast and accurate methods for estimating GFA, its bias, and standard deviation.
• To introduce deep, fully-connected neural networks for rapid dMRI metric estimation.
• To compare neural network performance against established statistical techniques.

Main Methods:

• Developed three deep, fully-connected neural networks for GFA, GFA bias, and GFA standard deviation estimation.
• Utilized Simulation Extrapolation (SIMEX) for bias estimation and wild bootstrap for variance estimation as comparison methods.
• Compared neural network predictions against observed GFA values and Q-ball model fits.

Main Results:

• The GFA neural network achieved higher accuracy than Q-ball fitting (RMSE 0.0077 vs. 0.0082).
• The bias estimation network significantly outperformed SIMEX (RMSE 0.0071 vs. 0.01).
• Neural networks provided faster and more accurate estimations of GFA and its associated error metrics.

Conclusions:

• Deep neural networks offer a computationally efficient and accurate alternative for dMRI quality assessment.
• This approach enhances the reliability and speed of analyzing diffusion MRI data.
• The developed methods hold promise for improving reproducibility in neuroimaging research.