CAAL-Unet: a Confusion Area Attention Lightweight Network for Brain Tumor Segmentation

Abstract

Brain tumors are one of the deadliest types of cancer. Accurate segmentation of brain tumors is very important for the diagnosis, evaluation and treatment of cancer. Recently, a large number of deep learning-based methods have replaced manual annotation and achieved remarkable results. However, most current research on CNNs is devoted to creating increasingly complex convolutional modules to improve performance, which in turn increases the computational cost and model complexity, which hinders its clinical application. This work proposes a effective CNN based on 3DUnet, CAAL-Unet. It uses 3D partical Convolution to replace the traditional convolution, thereby reducing network parameters and redundant feature maps. In the upsampling stage, a new confusion area attention module is added to enable the model to pay more attention to the edge part of the segmentation, to achieve multi-scale semantic feature information extraction of brain tumors. And the output of the higher level of the network is supervised to achieve deep injection of gradients. The performance and complexity of the model are evaluated on the Multimodal Brain Tumor Segmentation Challenge (BraTS2021) dataset. The results show that compared with other high-performance methods, our method has significantly reduced parameters and FLOPs, while maintaining competitive segmentation results.