A novel weighted probabilistic neural network for MR image segmentation

In this paper, we propose a modified probabilistic neural network, and applied it to brain magnetic resonance (MR) image segmentation. Comparing to conventional probabilistic neural network (PNN), the so-called weighted probabilistic neural network (WPNN): 1) uses covariance matrices instead of singular smoothing factor in kernel Junction; 2) adds weighting factors in pattern to summation layer. The WPNN is able to take partial volume effect, which exists commonly in MR imaging, into account not only in the final resulting stage, but also in the modeling process. We adopt self-organizing map (SOM) neural network as the previous step of WPNN classifier, to quantize the input data. SOM yields reference vectors necessary for probabilistic density function estimation. A supervised "soft" labeling mechanism based on Bayesian algorithm is developed, such that weighting factors can be generated for each SOM reference vector. Tissue segmentation results from various algorithms are compared, and the effectiveness and robustness of the proposed approach are demonstrated.