This study demonstrates the utility of a neural network classifier in separating significant cancer from low-grade cancer and non-cancerous lesions, based on the quantitative MRF and diffusion mapping. Using targeted biopsy data for training, the neural network classifier outperforms the linear regression model in both peripheral zone (PZ) and transition zone (TZ). The differentiation results showed an AUC of 0.90 in PZ and AUC of 0.89 in the TZ, comparing to AUC of 0.86 and 0.81 using Logistic Regression respectively. After applying the adaptive data oversampling algorithm, the AUC in characterizing TZ lesions can reach 0.96. Further classification utilizing patient clinical information showed statistically better accuracy in PZ while worse in TZ.