Automated Brain Tumor Segmentation

Abstract

The death rate and mortality of brain tumor are increasing these years and are seriously jeopardizing people’s life and health. Thus, the detection and segmentation of brain tumor are major challenge in its treatment planning. Magnetic Resonance Imaging (MRI) is a critical technique for precise detection of various pathologies, and it is widely used in medical imaging for the treatment of brain tumor. It is not easy to segment the brain tumors accurately due to its blurred boundary lines with surrounding tissues, complicated shapes and uneven gray scales. Manual segmentation is the most common way to achieve accurate segmentation. However, manual segmentation of brain tumor spends much time, and it is highly depended on radiologist’s experience. Therefore, automated brain tumor segmentation would have a big impact on brain tumor therapy and assessment. In this research, an automated segmentation model based on U-Net networks architecture [20] is proposed to accurately segment brain tumors. This method is evaluated on brain tumor data set shared by Dr. Jun Chen from School of Biomedical Engineering Southern Medical University on figshare, which contain 3064 T1-weighted brain tumor images from 233 patients. Comparing with classical U-Net networks, I added the Batch Normalization in U-Net architecture, cross-entropy and Adaptive Moment Estimation (Adam) [18] was used as loss function and optimizer respectively, which are specially designed for image segmentation tasks. The experimental data show that brain segmentation by U-Net convolutional network makes it possible to have more accurate result than Region Growing Method, and this model can achieve brain tumor segmentation without human intervention