Promptable diffusion-enhanced foundation framework for explainable brain tumor segmentation and classification

The diagnosis of brain tumors by magnetic resonance imaging (MRI) is still a difficult task because of the tumor heterogeneity, low tumor contrast, imaging artifacts and differences between tumor types. Deep learning methods commonly used today lack explainability, tolerability to corrupted imaging conditions and the ability to interpret the complex structures of tumors. To overcome these challenges, the study introduces a Promptable Diffusion-Enhanced Foundation Framework for Explainable Brain Tumor Segmentation and Classification with the recent BRISC 2025 dataset. The proposed framework leverages diffusion-based MRI enhancement, foundation-based feature representation learning, prompt-based attention mechanisms, and explainable dual-task intelligence to achieve simultaneous tumor segmentation and classification. A diffusion enhancement module is used to enhance the visibility of the tumor boundaries and to minimize the imaging artifacts before extracting features. Then, a pre-trained foundation encoder extracts strong semantic representations, and the adaptive localization of the tumor areas is achieved by prompt-guided attention. The framework jointly segments and classifies with an enhanced interpretation of the result with attention visualization and attention-based explanation maps. Experimental results showed that the classification accuracy of 98.1%, Dice score of 93.2%, and AUC of 99.0% were achieved, which is more superior than the conventional CNN and the transformer-based methods. The proposed framework also demonstrated high robustness in the presence of imaging noise and low MRI contrast conditions, which would be important for the clinical robust and interpretable analysis of brain tumors.