Training Data Distribution Significantly Impacts the Estimation of Tissue Microstructure with Machine Learning
PurposeSupervised machine learning (ML) provides a compelling alternative to traditional model fitting for parameter mapping in quantitative MRI. The aim of this work is to demonstrate and quantify the effect of different training strategies on the accuracy and precision of parameter estimates when supervised ML is used for fitting. MethodsWe fit a two-compartment biophysical model to diffusion measurements from in-vivo human brain, as well as simulated diffusion data, using both traditional model fitting and supervised ML. For supervised ML, we train several artificial neural networks, as well as random forest regressors, on different distributions of ground truth parameters. We compare the accuracy and precision of parameter estimates obtained from the different estimation approaches using synthetic test data. ResultsWhen the distribution of parameter combinations in the training set matches those observed in similar data sets, we observe high precision, but inaccurate estimates for atypical parameter combinations. In contrast, when training data is sampled uniformly from the entire plausible parameter space, estimates tend to be more accurate for atypical parameter combinations but may have lower precision for typical parameter combinations. ConclusionThis work highlights the need to consider the choice of training data when deploying supervised ML for estimating microstructural metrics, as performance depends strongly on the training-set distribution. We show that high precision obtained using ML may mask strong bias, and visual assessment of the parameter maps is not sufficient for evaluating the quality of the estimates.