Optimizing Convolutional Neural Networks for Brain Tumor Segmentation in MRI Images

Abstract

Gliomas comprise about 80% of all malignant brain tumors and have the lowest survival rate of all brain tumors. Segmentation of tumors is an important step in evaluating the tumor, preparing the treatment plan and estimating the patient survival period. Tumor tissues have a distinguishable appearance in MRI images so they are widely used for brain tumor segmentation. Many solutions were proposed to automate brain tumor segmentation but convolutional neural networks (CNNs) have the most promising results. Tens of neural networks were proposed for tumor segmentation but they still did not achieve good enough accuracies to be deployed in real-world applications. In this paper, we focused on optimizing patch-wise classifier CNN and the results obtained are discussed to show the effect of some design decision taken. We evaluated the segmentation results using the Dice Similarity Coefficient (DSC). The results of this paper can be used to improve existing models or as a guideline for developing new CNN models. Finally, we point out possible future directions for research.