Enhanced Brain MRI Analysis with Explainable AI and Data Fusion

Accurate and interpretable brain tumor diagnosis remains a critical challenge in neuro-oncology due to the complexity of multimodal MRI data and the black-box nature of many AI models. This thesis proposes a novel deep learning pipeline that combines multimodal MRI fusion with dual-layer explainable AI (XAI) to enhance both segmentation and classification of gliomas. The framework integrates T1, T1Gd, T2, and FLAIR modalities using feature-level fusion and employs SHAP and Grad-CAM to generate interpretable outputs for clinicians. Extensive evaluation on the BraTS dataset demonstrated superior performance: Dice scores of 0.946 (Whole Tumor), 0.968 (Tumor Core), and 0.975 (Enhancing Tumor), and a glioma classification AUC of 0.964. GAN-based synthesis restored up to 90% performance when modalities were missing. The model achieved fast inference (2.3 seconds/scan) with a lightweight architecture (28.7M parameters). Radiologist assessments rated the explanation clarity at 4.7/5 on average, supporting clinical trust. The proposed system balances performance, transparency, and efficiency, making it suitable for real-world integration into diagnostic workflows. Its robustness under data variability and integration of interpretability mechanisms represents a significant advancement toward clinically viable AI tools in medical imaging.