Ludovico Minati1. 1. Fondazione Istituto Nazionale Neurologico Carlo Besta, via Celoria 11, Milano, Italy. lminati@istituto-besta.it
Abstract
OBJECT: To investigate whether multi-layer perceptrons (MLPs) could be used to determine biexponential and diffusional kurtosis model parameters directly from diffusion-weighted images. MATERIALS AND METHODS: Model parameters were determined with least-squares fitting and with MLPs. The corresponding estimates were compared with linear regressions, t tests and Levene's tests. Residuals were also compared. RESULTS: Strong linear correlation was found for all parameters. MLP estimates were unbiased for the biexponential but not for the kurtosis model, and generally had smaller variance. Residuals were smaller for MLP estimates. The maps generated by the two methods were visually very similar. CONCLUSION: Multi-layer perceptrons are potentially useful as a curve fitting method for these models.
OBJECT: To investigate whether multi-layer perceptrons (MLPs) could be used to determine biexponential and diffusional kurtosis model parameters directly from diffusion-weighted images. MATERIALS AND METHODS: Model parameters were determined with least-squares fitting and with MLPs. The corresponding estimates were compared with linear regressions, t tests and Levene's tests. Residuals were also compared. RESULTS: Strong linear correlation was found for all parameters. MLP estimates were unbiased for the biexponential but not for the kurtosis model, and generally had smaller variance. Residuals were smaller for MLP estimates. The maps generated by the two methods were visually very similar. CONCLUSION: Multi-layer perceptrons are potentially useful as a curve fitting method for these models.