An Optimized Segmentation Framework Applied to Glioma Delimitation

In this article we describe our segmentation framework applied to glioma delimitation in multimodal magnetic resonance images. Statistical pattern recognition strategies are applied to create a discriminative function. The discriminative classifier is the result of an automatic learning process based on random forest (RF) algorithm. This algorithm is used for two different purposes as well as in the construction of segmentation classifiers, as in the variable importance evaluation task. In the training phase the most important local image features are selected and the most adequate optimal parameters of the RF classifier are determined. The goal is to find the discriminative model that allows us to obtain the best possible segmentation performances. The segmentation framework obtained was evaluated online using the brain tumor segmentation benchmark system, and the performances were compared to the best ones reported in the literature.