Noise Level Matching Improves Robustness of Diffusion Mri Parameter Inference by Synthetic Q-Space Learning

Abstract

Diffusion MRI is a powerful tool for characterizing the local properties of microstructures in living organisms, using parameters of various signal models. A diffusion MRI dataset consists of signals measured using a variety of directions and strengths of the gradient field expressed in q-space. Machine learning approaches called q-space learning have been recently employed to infer the parameters of diffusion MRI using deep regression neural networks, etc., instead of conventional model parameter fitting. Moreover, a training approach that uses synthetically generated data can overcome the limitations of training on real MRI data. During the generation of synthetic training data, signal values in q-space can be contaminated by artificial noise for realistic data synthesis. In this study, we experimentally show that the amount of noise between the training and test data should be matched to obtain optimal robustness. As examples of the signal model, diffusion tensor imaging, diffusional kurtosis imaging, and neurite orientation dispersion diffusion imaging are used in experiments with synthetic and empirical datasets.