An Explainability-Aware Transformer Framework for Brain Tumor Segmentation and Classification Using MRI

Magnetic Resonance Imaging is one of the most commonly used neuro-oncology imaging modalities, which is a non-invasive mode of imaging and helps in detecting brain abnormalities in an effective way. Earlier researchers have demonstrated that brain tumor segmentation and classification can be effectively performed using deep learning techniques. Existing studies are primarily aimed at increasing prediction accuracy and provide insignificant consideration to model interpretability, limiting their practical application in clinical practice. To address this limitation, this research presents a two-stage explainable deep learning model, which combines transformer-based segmentation with an ensemble classification model that is consistent in explanations. The first stage introduces Swin-DS-HAFUNetv2, an enhanced transformer-based segmentation architecture that integrates hierarchical Swin Transformer encoders, refined hierarchical attention fusion, a contextual bottleneck transformer, and multi-scale deep supervision to improve tumor localization in T1-weighted MRI, particularly under low-contrast and irregular morphological conditions. The second stage includes the ECWMEv2 ensemble classifier, which integrates a perturbation analysis based on Grad-CAM (Gradient-weighted Class Activation Mapping) to systematically assess the consistency and clinical significance of visual explanations for candidate models. Only those architectures that exhibit stable and pathology-consistent explanations, such as ConvNeXt, Swin Transformer, and EVA02, are stored and merged by means of explanation-weighted soft voting with XGBoost-based meta-learning. Experimental evaluation on the BRISC2025 benchmark dataset indicates that Swin-DS-HAFUNetv2 has a mean Dice coefficient of 0.9782 and Intersection over Union (IoU) of 0.8656, with ECWMEv2 having a classification accuracy of 0.9917 and a Macro F1 score of 0.9867. The mean Grad-CAM IoU of 0.692 reflects uniform and anatomically significant consistency of attention to tumor regions. These results demonstrate that the integration of explanation stability as a fundamental design principle significantly improves model robustness and interpretability to provide a methodologically validated and benchmark-level framework for future studies of multimodal and clinically oriented brain tumor analysis systems.