Parameter estimation for WMTI-Watson model of white matter using encoder-decoder recurrent neural network

Area of Science:

• Neuroimaging
• Biophysics
• Computational Neuroscience

Background:

• Diffusion MRI (dMRI) biophysical modeling estimates tissue microstructure.
• Non-linear least squares (NLLS) is common but computationally expensive and prone to local minima.
• Deep learning models require retraining for different acquisition protocols and noise levels.

Purpose of the Study:

• To develop a novel, computationally efficient, and generalizable fitting approach for dMRI modeling.
• To introduce an encoder-decoder recurrent neural network (RNN) solver for accelerated model estimation.
• To improve upon the limitations of NLLS and existing deep learning methods in dMRI analysis.

Main Methods:

• The study proposed an encoder-decoder recurrent neural network (RNN) solver.
• The RNN solver estimates the white matter tract integrity (WMTI)-Watson model from diffusion and kurtosis imaging (DKI) parameters.
• Methods were evaluated on synthetic and in vivo rat and human brain datasets, comparing RNN against analytical solutions, NLLS, and multilayer perceptron (MLP).

Main Results:

• The RNN solver significantly reduced computation time compared to analytical solutions and NLLS.
• The RNN approach demonstrated comparable accuracy and improved robustness.
• Superior generalizability across different datasets was observed for the RNN method compared to MLP.

Conclusions:

• The proposed RNN estimator offers a faster and more robust alternative for dMRI model fitting.
• Its ability to generalize across various datasets without retraining highlights its potential for widespread clinical and research use.
• This method advances the field of neuroimaging by providing a more efficient tool for microstructural tissue property estimation.