Automatic Segmentation on Glioblastoma Brain Tumor Magnetic Resonance Imaging Using Modified U-Net

Abstract

Glioblastoma is listed as a malignant brain tumor. Due to its heterogeneous composition in one area of the tumor, the area of tumor is difficult to segment from healthy tissue. On the other side, the segmentation of brain tumor MRI imaging is also erroneous and takes time because of the large MRI image data. An automated segmentation approach based on fully convolutional architecture was developed to overcome the problem. One of fully convolutional network that used is U-Net framework. U-Net architecture is evaluated base on the number of epochs and drop-out values to achieve the most suitable architecture for the automatic segmentation of glioblastoma brain tumors. Through experimental findings, the most fitting architectural model is mU-Net architecture with an epoch number of 90 and a drop out layer value of 0.5. The results of the segmentation performance are shown by a dice value of 0.909 which is greater than that of the previous research.

Keywords

• Brain tumor segmentation
• fully convolutional network
• U-Net
• drop-out
• dice score