Prostate Magnetic Resonance Imaging (MRI) is one of the most promising approaches to facilitate prostate cancer diagnosis. The effort of research community is focused on classification techniques of MR images in order to predict the cancer position and its aggressiveness. The reduction of False Negatives (FNs) is a key aspect to reduce mispredictions and to increase sensitivity. In order to deal with this issue, the most common approaches add extra filtering algorithms after the classification step; unfortunately, this solution increases the prediction time and it may introduce errors. The aim of this study is to present a methodology implementing a 3D voxel-wise neighborhood features evaluation within a Support Vector Machine (SVM) classification model. When compared with a common single-voxel-wise classification, the presented technique increases both specificity and sensitivity of the classifier, without impacting on its performances. Different neighborhood sizes have been tested to prove the overall good performance of the classification.