EFFICIENT BRAIN TUMOR SEGMENTATION WITH 3D U-NET AND PRETRAINED BACKBONES

Abstract

Segmentation of a brain tumor from three-dimensional data is one of the very important and very challenging tasks within the scope of medical image analysis since whatever is manual-segmented by human operators may lead to imprecision in diagnosis and treatment. Some new methods have recently been developed for tumor segmentation within magnetic resonance imaging data using two-dimensional and three-dimensional convolutions. However, a 2D convolution does not make full user of the spatial information which comes natural with volumetric study medical imaging data, and 3D convolutions come with much computation overhead and much memory resource. This tries to address that problem by using the 3D U-Net architecture, specifically for brain tumor MRI images. Effects of the use of pre-trained backbone networks into the 3D U-Net model were studied here within three main configurations: the usual 3D U-Net, 3D U-Net with enhancement using ResNet50, and 3D U-Net incorporated with VGG16. The model was tested with actual data obtained from the BraTS2020 datasets. A compare and validate analysis were done. Different metrics like accuracy, Dice Coefficient, cross-entropy loss, precision, sensitivity, and specificity were focused on in the comparison. The accuracies of 3D UNet, 3D-UNet pre-trained with ResNet50 backbone, and 3D-UNet pre-trained with VGG16 backbone were found to be 0.9798, 0.9795, and 0.9810, respectively. Dice Coefficient scores were noted as 0.9993, 0.9976, and 0.9996 in the same order. Therefore, these proposed methods are competitive with state-of-the-art approaches in this specific domain. But better results were found with the model that combines the 3D UNet model with a VGG16 pre-trained backbone for segmentation of brain tumors.