Brain Tumor Segmentation Using U-Net and Edge Contour Enhancement

Abstract

Segmentation of brain tumors by magnetic resonance imaging (MRI) plays a pivotal role in evaluating the disease condition and deciding on a future treatment plan. This type of segmentation task usually requires extensive experience from medical practitioners and enormous amounts of time. To mitigate these issues, this study deploys a segmentation model for brain tumors based on U-Net and a comprehensive data processing approach, including target magnification and image transformation, such as data augmentation and edge contour enhancement. Compared with the manual segmentation of radiologists, which is considered the gold standard, the proposed model revealed good performance and yielded a median dice similarity coefficient of 0.637 (interquartile range: 0.382-0.803) for brain tumor segmentation. Results with and without edge contour enhancement demonstrated significant differences based on the Wilcoxon signed-tank test with P = 0.028. The proposed model enables effective segmentation of brain tumors determined by MRI and can assist medical practitioners tasked with analyzing complicated medical images.