The brain is the center of human control and communication. Hence, it is very important to protect it and provide ideal conditions for it to function. Brain cancer remains one of the leading causes of death in the world, and the detection of malignant brain tumors is a priority in medical image segmentation. The brain tumor segmentation task aims to identify the pixels that belong to the abnormal areas when compared to normal tissue. Deep learning has shown in recent years its power to solve this problem, especially the U-Net-like architectures. In this paper, we proposed an efficient U-Net architecture with three different encoders: VGG-19, ResNet50, and MobileNetV2. This is based on transfer learning followed by a bidirectional features pyramid network applied to each encoder to obtain more spatial pertinent features. Then, we fused the feature maps extracted from the output of each network and merged them into our decoder with an attention mechanism. The method was evaluated on the BraTS 2020 dataset to segment the different types of tumors and the results show a good performance in terms of dice similarity, with coefficients of 0.8741, 0.8069, and 0.7033 for the whole tumor, core tumor, and enhancing tumor, respectively.
Keywords: U-Net; attention; brain tumor segmentation; deep learning; encoder; pyramid neural network; transfer learning.