Machine learning for improved diffusion MRI parameter estimation with gradient nonlinearity correction

Abstract

White matter (WM) degradation is one of the most common lesions causing neurological disorders, where early diagnosis is crucial. Diffusion tensor imaging is a widely used method for the study of WM injuries in the brain, however its estimation using traditional methods such as non-linear least squares (NLLS) is time consuming and patient dependent. This work presents a machine learning (ML) approach for diffusion MRI parameter estimation. A feed forward network following two different strategies is presented to attempt the fitting of a symmetric diffusion tensor model. The possibility of realtime mapping is approached by four different experimental setups with increasing data complexity during the ML training, including synthetic and real signal simulations. Moreover, a novel approach on gradient nonlinearity (GNL) correction using ML is presented, opening the possibility of correcting for spatially varying b-values and b-vectors while training the model. The results showed an acceptable ML performance compared to NLLS when training and testing with one single subject, demonstrating that ML can be used for parameter estimation of diffusion images and GNL correction. However, the generalization of the network to accept more than one subject is still a challenge. Further hyperparameter tuning and architecture configuration experiments are needed to generate comparable results to NLLS. Nevertheless, these initial results highlighted crucial aspects in the fitting process that could be important for future research of the topic.