ViTGradX: An Explainable Vision Transformer Framework and Deep Survival Modeling for multimodal Brain Tumor classification and Prognosis

Brain tumors require timely diagnosis and accurate prognosis for efficient treatment. Magnetic Resonance Imaging (MRI) is the standard for its non-invasive nature and high-resolution imaging. While deep learning (DL) approaches have shown promise in brain tumor segmentation and arrangement, they often lack interpretability and comprehensive survival prediction, limiting their clinical adoption. To address these issues, ViTGradX, an explainable Vision Transformer (ViT) framework integrated with Grad-CAM, is proposed for multimodal brain tumor analysis. This system performs tumor detection, classification, and survival prediction. The pipeline incorporates preprocessing steps like contrast enhancement and skull stripping to improve image clarity. Tumor segmentation is carried out using a Structure Correcting Adversarial Network (SCAN), and hybrid structures are mined by combining Binary Robust Independent Elementary Features (BRIEF) and Discrete Wavelet Transform (DWT) to capture both local and frequency-based characteristics. ViTGradX first performs binary classification to distinguish between normal and abnormal cases, then categorizes them into non-enhancing cores (NET), necrotic (NCR), peritumoral edema (ED), and GD-enhancing tumors (ET). Grad-CAM enhances interpretability with visual attention maps, while DeepSurv estimates survival risk scores. The model achieves excellent performance with 97.67% recall, 96.45% F1-score, 99.99% accuracy, and 98.70% precision, showing strong potential as an explainable clinical decision-support system.