Brain tumor segmentation with corner attention and high-dimensional perceptual loss

Abstract

Accurate segmentation of brain tumors in MRI sequences is an essential factor that helps doctors make detailed surgery plans and evaluate prognoses. However, due to the diversity of tumors and the complexity of subregions, accurate brain tumor segmentation is still a major challenge. The inter-slice dimension is important in MRI scans as it characterizes the changes in tumors and enriches the contextual information, which can help networks to predict more finer segmentation results. To extract the inter-slice dimension information of volume data and capture rich context dependence, we propose a corner attention module (CAM), which can effectively model the relationship between the sagittal axis, coronal axis, and axial axis, as well as extract complementary information between inter-slices and intra-slices. Furthermore, empowered by the novel high-dimensional perceptual loss (HDPL), the model can preserve local consistency and explore perceptual similarity, which makes predictions and ground-truth similar in high-dimensional space and the boundary of prediction finer. Combining CAM and HDPL on the basis of U-Net, we propose CH-UNet for brain tumor segmentation. Extensive experiments on the authoritative public benchmarks BraTs2018, BraTs2019, and BraTs2020 reveal that our approach presents an average improvement of 3.01% on the dice coefficient and an average decreasement of 42.27% on 95th Hausdorff distance compared to the baseline. In addition, compared with state-of-the-art methods, our approach exhibits competitive segmentation performance and has the potential to be implemented in clinical medical applications.