Brain tumor segmentation from MRI images using convolutional neural networks

Abstract

Brain tumors result from the accumulation of abnormal cells inside the brain. The process of brain tumor segmentation plays a vital role in detecting early-stage brain tumors. There are several challenges in the tumor segmentation process due to the variations in size, location, and intensity of brain tissues. Traditional brain tumor segmentation methods are very time-consuming as segmentation is carried out manually. Automated segmentation methods are necessary for rapid diagnosis and treatment. In our paper, we highlight the complications surrounding a brain tumor and use an encoder-decoder-based approach of CNN algorithms to train segmentation models that will help identify and localize brain tumors with the utmost accuracy. We have used four CNN architectures to train our models, namely UNet, ResNet50, ResNext50, and EfficientNetB7. For our encoder-decoder models, we used ResNet50, ResNext50, and EfficientNetB7 as the encoder blocks of U-Net in three different models. Hence, we performed three experiments using these four architectures and compared their performance. We obtained the best results using the U-Net + EfficientNetB7 model.