MGAEPL: Multi-Granularity Automated and Editable Prompt Learning for brain tumor segmentation

Brain tumor segmentation is a critical task for accurate diagnosis, treatment planning, and patient prognosis. However, it remains challenging due to the small size and irregular morphology of tumor subregions. Existing approaches often struggle to incorporate valuable medical prior knowledge, such as tumor grade, which is closely associated with tumor heterogeneity and biological aggressiveness. To address this limitation, we propose a novel framework named MGAEPL (Multi-Granularity Automated and Editable Prompt Learning). MGAEPL employs a shared encoder to extract imaging features for simultaneous tumor-grade prediction and segmentation mask generation. The predicted tumor grade is then utilized as an automatically generated prompt, which is processed by a multi-granularity prompt encoder to guide the decoder in producing accurate segmentation masks. This design enables the integration of medical priors into the segmentation process without requiring manual prompt input. Extensive experiments on the BraTS 2018 and BraTS 2021 datasets demonstrate the effectiveness and generalizability of MGAEPL, highlighting its potential for clinical deployment in brain glioma segmentation tasks.