ONCOhabitats is an open online service that provides a fully automatic analysis of tumor vascular heterogeneity in gliomas based on multiparametric MRI. Having a model capable of accurately segment pathological tissues is critical to generate a robust analysis of vascular heterogeneity. In this study we present the segmentation model embedded in ONCOhabitats and its performance obtained on the BRATS 2019 dataset. The model implements an residual-Inception U-Net convolutional neural network, incorporating several pre- and post- processing stages. A relabeling strategy has been applied to improve the segmentation of the necrosis of high-grade gliomas and the non-enhancing tumor of low-grade gliomas. The model was trained using 335 cases from the BraTS 2019 challenge training dataset and evaluated with 125 cases from the validation set and 166 cases from the test set. The results on the validation dataset in terms of the mean/median Dice coefficient are 0.73/0.85 in the enhancing tumor region, 0.90/0.92 in the whole tumor, and 0.78/0.89 in the tumor core. The Dice results obtained in the independent test are 0.78/0.84, 0.88/0.92 and 0.83/0.92 respectively for the same sub-compartments of the lesion.