A brain tumor segmentation method based on attention mechanism

Abstract With the exacerbation of global population aging, the incidence of brain tumors continues to rise, placing unprecedented pressure on medical institutions in their diagnostic and treatment work. Therefore, in clinical practice, providing precise and reliable auxiliary diagnostic information to physicians has become crucial to meet the demand for efficient and accurate diagnostics. In order to assist neurosurgeons in accurate diagnosis and surgical planning on brain MRI data, this study proposes a novel attention-based brain tumor segmentation algorithm, named BSAU-Net. Considering the complexity and diversity of brain tumor edges, this study introduces an Edge Aware (EA) module to enhance the sensitivity of the BSAU-Net to tumor regions while preserving the integrity of their edge contours. The importance of global spatial relationships of tumors for precise segmentation is acknowledged by proposing a Spatial Self-Attention (SPA) module, which establishes correlations between global features, thereby improving the model's segmentation performance. To address the issue of class imbalance, a novel loss function, BADLoss, is designed to alleviate the performance degradation of BSAU-Net. The BSAU-Net algorithm, as proposed, demonstrates exceptional performance on the BraTS2018 and BraTS2021 datasets, with an average Dice coefficient of 0.7506, Positive Predictive Value (PPV) of 0.7863, and Sensitivity of 0.8998.