Segmenting Brain Tumor with an Improved U-Net Architecture

Abstract

To aid in the clinical diagnosis of brain tumors, magnetic resonance imaging (MRI) is frequently used. The amount of time it takes to manually segment MRI images depends on the radiologist’s level of expertise. In this paper, a novel U-Net architecture for segmenting images of brain tumors is proposed. We have evaluated BraTS 2020 dataset with an improved U-Net structure with a dropout layer inserted between the encoder and decoder to reduce overfitting. By comparing with other U-Net architectures, our method has shown a promising result with dice coefficients 70.40%, 69.08% and 73.03%, for whole tumor (WT), tumor core (TC) and enhanced tumor (ET).