Automatic Brain Tumor Segmentation Using Multi-scale Features and Attention Mechanism

Gliomas are the most common primary malignant tumors of the brain. Magnetic resonance (MR) imaging is one of the main detection methods of brain tumors, so accurate segmentation of brain tumors from MR images has important clinical significance in the whole process of diagnosis. At present, most popular automatic medical image segmentation methods are based on deep learning. Many researchers have developed convolutional neural network and applied it to brain tumor segmentation, and proved superior performance. In this paper, we propose a novel deep learned-based method named multi-scale feature recalibration network(MSFR-Net), which can extract features with multiple scales and recalibrate them through the multi-scale feature extraction and recalibration (MSFER) module. In addition, we improve the segmentation performance by exploiting cross-entropy and dice loss to solve the class imbalance problem. We evaluate our proposed architecture on the brain tumor segmentation challenges (BraTS) 2021 test dataset. The proposed method achieved 89.15%, 83.02%, 82.08% dice coefficients for the whole tumor, tumor core and enhancing tumor, respectively.