Impact of Training Set Sampling on Parameter Estimation Neural Networks for Monoexponential, Biexponential, and mcDESPOT Models with applications in Myelin Water Imaging

Parameter estimation neural networks (PENNs) are increasingly used for quantitative myelin water imaging, where accurate mapping of biophysical parameters is essential. PENNs are trained on synthetic datasets composed of signals paired with known parameters. However, for a given parameter range, the construction of the training set has for PENN has not been extensively studied in spite of its impact on computational time, and, potentially, accuracy. For example, while finer discretization would be expected to improve parameter estimation accuracy, it also increases computational cost, especially for multidimensional parameter spaces. In this study, we systematically investigate the effect of training set discretization and distribution on PENN accuracy across monoexponential, biexponential, and mcDESPOT signal models. Results show that for the biexponential and mcDESPOT model, accuracy improves asymptotically with increasing discretization, beyond which further refinement offers negligible benefit, whereas the monoexponential model achieves high accuracy even with sparse training sets. In addition, the distribution of training set samples in parameter space is also shown to have a substantial impact on NN performance. These results provide practical guidance for optimizing training set design in neural network–based parameter estimation for myelin water imaging and related biophysical modeling applications.