Automatic segmentation of glioblastoma multiforme and its peripheric edema using Otsu, Chan-Vese and a neural network

Abstract

Glioblastoma multiforme is the primary brain tumor most agressive and of worst prognosis. At present, the automatic segmentation of this kind of tumor is being intensively studied, as it has relevant utilities related to diagnosis and prognosis. In this work, an automatic segmentation is achieved in base on the four basic MRI modalities and a combined algorithm that articulates classic image-processing methods and a multilayer neural network. The network is feed by 30 selected intensity and texture features that classify each pixel in one of four classes. The complete algorithm achieve Dice coefficients of 89%, 81%, 80%, 66% and 84% for the whole tumor area, contrast-enhancing tumor, edema, necrosis and tumor core segmentations, respectively. These coefficients are in the range of the better obtained in the literature.