Segmentation of Brain Glioma in MRI Images Using Deep Learning

A very common form of brain cancer is glioma which is the tumor originates from the glial cells in the brain or the spine. It makes up for about 80% of brain malignancies altogether, with the highest deaths recorded. Magnetic resonance imaging (MRI) is the typical imaging type to capture the structure and the physiology of the brain, including in diagnosing the gliomas. Other than that, these images are used in managing the treatment and for surgical purposes. For these procedures, the accurate boundary of the glioma needs to be determined through segmentation. This task is normally performed by the expert. To reduce bias, inter-expert variability and the time spent, automated software-based segmentation tool is applied. This paper presents the segmentation method of the gliomas in MRI images using a using a deep learning-based model architecture segmentation method, which is U-Net. The performance of U-Net is tested on BRATS datasets that consists of patients' data together with the ground truth data. Experimental results on the BraTS2018, BraTS2019 and BraTS2020 validation dataset demonstrates that the proposed model achieved average Dice Scores of 0.97, 0.97, 0.96 for whole tumor, 0.97, 0.98, 0.95 for tumor core and 0.95, 0.94, 0.95 for enhancing tumor respectively. In conclusion, studies conducted using U-Net methods have obtained the image's results of brain glioma segmentation that show the accuracy almost identical to the ground truth segmentation images.