An efficient deep learning approach for brain tumor detection using 3D convolutional neural network

Abstract

In medical application, deep learning-based biomedical pixel-wise detection through semantic segmentation has provided excellent results and proven to be efficient than manual segmentation by human interaction in various cases. A well-known and widely used architecture for biomedical segmentation is U-Net. In this work, a convolutional neural architecture based on 3D U-Net but with fewer parameters and lower computational cost is used for pixel-level detection of brain tumor through semantic segmentation. The proposed model is able to maintain a very efficient performance and provides better results in some cases compared to conventional U Net, while reducing memory usage, training time and inference time. BraTS 2021 dataset is used to evaluate the proposed architecture and it is able to achieve Dice scores of 0.9105 on Whole Tumor(WT), 0.884 on Tumor Core(TC) and 0.8254 on Enhancing-Tumor(ET) on the testing dataset.