Segmentation of low-grade gliomas using U-Net VGG16 with transfer learning

Around 2000 cases of gliomas are diagnosed every year in the US, representing 23.41 percent of all primary brain tumors. World Health Organization (WHO) grade II gliomas or Low-Grade Gliomas (LGG) are slow-growing brain tumors. LGG is a fatal disease of young adults (between 35 and 44 years of age). LGG can transform into High-Grade Gliomas (HGG) or WHO grades III and IV occurred in most patients and ultimately leading to death. General treatment for LGG patients is surgical resection, radiotherapy, and chemotherapy. Fluid-Attenuated Inversion Recovery (FLAIR) imaging is needed to determine the tumor location before doing surgical resection. We propose a combined architectural innovation of U-Net and VGG16 with transfer learning as a hybrid model for tumor segmentation. Employing the preoperative FLAIR imaging data of 110 patients with LGG from the Cancer Genome Atlas, this deep learning algorithm achieves a high result with the Dice Similarity Coefficient of 99% and the Area Under Curve (AUC) of 98%, better than the previous approach done by Buda, et al.