TAE-ResU-Net:A trinomial attention-based channel-interactive U-shaped residual network
Yu Chen, Rui Wang, Tian Gao, Shuo Zhang · 2022
Brain MRI images are characterized by complex data and large data volume, and accurate and reliable segmentation of brain MRI tumor images is a challenge due to the variability of lesion size, shape, and location. In recent years, many researchers have used deep learning-based convolutional neural networks for image segmentation of medical MRI, CT, etc., and have made great progress and achievements in the field of segmentation. However, the convolutional neural network-based model is limited by the local Receptive field, which cannot capture the relevant information at a long distance, and cannot achieve fine segmentation of brain lesion object boundaries, and the performance of small object segmentation is poor. For this reason, some researchers have proposed a self-attention-based segmentation method, which improves the accuracy but has a huge number of parameters, long running time, and cannot be used on small devices, which has far-reaching effects on patient medical images. To this end, to address the above problems, this paper proposes a trinomial attention-based channel interaction residual U-shaped network to improve segmentation accuracy while reducing the number of model parameters as much as possible to achieve a balance between the number of model parameters and segmentation accuracy. We use the datasets of BraTS2018 and BraTS2019 for experiments, and the experimental results show that our model outperforms other more advanced models in most indicators such as DISC, Hausdorff distance, etc., and has certain advantages in tumor segmentation.