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

Özet

In prostate cancer detection and diagnosis, high b-value diffusion MR signals of the prostate tissues are of great concern. MR scanners and computation methods facedifficultiesin obtaining these signals. This study aims development of a neural network model to predict the MR signal amplitudes at high b-values from the amplitudes at low b-values. Synthetic diffusion MR signals are generated using a kurtosis model for noise-free and noisy conditionsconsidering nine b-values: the low b-valuesare 0, 50, 250, 500, 750s/mm2and the high b-values are 1000, 1250, 1500, 2000s/mm2. Four radial basis functionsneural networks (RBF-NN) connected in parallelare designed to accept the signal amplitudesat low b-values and to provide signal amplitudesat thehigh b-values.RBF-NNs housing altered number of neurons with radial basis functions attributing different widthsin the hidden layersof the networks are analyzed. Learning and prediction performances of the NNsare assessed from training and testing datasets. For the noise-free condition, RBF-NNsreveal perfect predictions(r= 1.000) withvery good learnings(MSE= 0.76-0.02×10-6). For the noisy conditions, the RBF-NNs achievemoderate to strong predictions(r= 0.981-0.463)withgood learnings (MSE= 0.32-10.33×10-3). Prediction performance reduces as the level of noise and/or targetedhigh b-value increases.RBF-NNsfacilitate prediction of high b-value diffusion MR signals of the prostate byrequiringno diffusion signal decay function, optimization algorithm or initial/boundary values for the optimization algorithm. They may be quite functional in accuratevoxel-wise generation of high b-value MR images for early detection and diagnosis of prostate cancer. Further prospective studies are needed tojustify the potentialbenefitsin clinicalpractice.