Glioma segmentation based on deep CNN

Brain tumor segmentation represents a hard job for radiologists as the brain is the most complicated and complex organ. Among the several brain tumors that existed, gliomas are the most aggressive and common. They lead to a short life in their highest grade especially. It is usually the most found tumors, which have various shapes, sizes, and brightness. It can appear anywhere in the brain. These causes make the automatic brain tumor segmentation a challenging problem. In this area, different Deep Learning (DL) models are suggested to help doctors. In this work, a new deep Convolutional Neural Network (CNN) architecture is presented to surpass these drawbacks. Our contributions incorporate three aspects. First, we exploited a pre-processing step based on intensity normalization with the goal to enhance the quality of the images. Second, we suggested an automatic segmentation model using CNN. The new scheme contains various convolutional layers, all exploiting 3 × 3 kernels, and one fully connected layer. Finally, we exploit a post-processing approach with the goal to ameliorate the segmentation results of the suggested model. We have evaluated the proposed technique based on the Multimodal Brain Tumor Image Segmentation Challenge datasets (BRATS 2017). The gained results provide the effectiveness of the suggested model compared with several techniques.

I. Introduction

Gliomas represent the brain tumors that having the highest prevalence and mortality rate [1]. They represent about 80% of malignant brain tumors [2]. They can be grouped into Low Grade Gliomas (LGG) and High Grade Gliomas (HGG), where the HGG is more aggressive and infiltrative. On average, the patients cannot survive more than 14 months after diagnosis, even under treatment.