Development of a Radial Basis Function Neural Network Model to Predict High b-value Diffusion MR Signals of the Prostate

Abstract

This study aims development of a neural network model to predict the signal amplitudes of diffusion MR signals of prostate tissues at high b-values from the amplitudes of low b-values obtained by using diffusion weighted imaging. Synthetic diffusion MR signals are generated using a kurtosis model for noise-free and noisy conditions considering nine b-values: the low b-values are 0, 50, 250, 500, 750s/mm² and the high b-values are 1000, 1250, 1500, 2000s/mm². Four radial basis functions neural networks (RBF-NN) connected in parallel are designed to accept the signal amplitudes at low b-values and to provide signal amplitudes at the high b-values. RBF-NNs housing altered number of neurons with radial basis functions attributing different widths in the hidden layers of the networks are analyzed. Learning and prediction performances of the NNs are assessed from training and testing datasets. For the noise-free condition, RBF-NNs reveal perfect predictions (r= 1.000) with very good learnings (MSE= 0.76-0.02×10^-6). For the noisy conditions, the RBF-NNs achieve moderate to strong predictions (r= 0.981-0.463) with good learnings (MSE= 0.32-10.33×10^-3). Prediction performance reduces as the level of noise and/or targeted high b-value increases. RBF-NNs facilitate prediction of high b-value diffusion MR signals of the prostate by requiring no diffusion signal decay function, optimization algorithm or initial/boundary values for the optimization algorithm. They may be quite functional in accurate voxel-wise generation of high b-value MR images for early detection and diagnosis of prostate cancer. Further prospective studies are needed to justify the potential benefits in clinical practice.