Hierarchical Cascaded Multi-Axis Window Self-Attention and Layer Feature Fusion for Brain Glioma Segmentation

Automatic segmentation of brain glioma from magnetic resonance imaging (MRI) using deep learning methods is very important for clinical diagnosis and follow-up treatments. The size, shape and location of glioma vary greatly among different patients, and the spatial information and location details of its corresponding feature map are very easy to lose, resulting in the segmentation accuracy of related models still to be improved. This paper proposes an advanced brain glioma segmentation network LHA-UNet based on hierarchical cascaded multi-axis window self-attention and multi-layer feature fusion. By comparing with some models such as U-Net, HyResUNet, DenseUNet, UNet++, LHA-UNet model has achieved the best results in all aspects, which proves the effectiveness of components of hierarchical cascaded multi-axis window self-attention and layer feature fusion.