Accurate segmentation of brain tumor images is crucial in medical auxiliary diagnosis. However, the complex morphology and ambiguous boundary contours of brain tumors pose significant challenges to precise segmentation.

To address these issues, we developed MISA-Net, which is based on enhanced multi-scale feature interactions and selective feature fusion attention. Initially, a Multi-Scale Feature Interaction (MSFI) module was implemented to enhance the interaction between features at different scales, resolving issues of misclassification in regions with complex tumor morphologies. Subsequently, a Selective Feature Fusion Attention (SFFA) mechanism was introduced to reduce the interference of redundant information in skip connections on crucial features.

Experiments on the BraTS 2019 dataset show that MISA-Net achieved Dice coefficients of 80.02%, 88.86%, and 86.02% in the enhancing, core, and whole tumor areas, respectively. Additionally, the Dice coefficient for the whole tumor area impressively reached 90.33% on the Kaggle LGG dataset; the Dice coefficient for the whole tumor area impressively reached 84.97% on the Figshare dataset.

Compared to existing mainstream models, MISA-Net demonstrates superior performance in brain tumor segmentation tasks, highlighting its potential and advantages in clinical diagnosis and treatment.