The correct diagnosis and prediction of malignancy in brain tumors are critical for neuro-oncology, as they directly influence clinical decision-making. Although deep learning models have had notable success in tumor classification and segmentation based on MRI data, most existing approaches are limited in three aspects: building on imaging modalities only, disregarding clinically relevant metadata, and lacking interpretability because of non-integrated explainable AI (XAI). To overcome these limitations, we present NeuroExplainAI, an explainable deep learning framework for a holistic brain tumor diagnosis and grading. We present NeuroFusionNet, a dual task architecture to fuse deep CNN features with hand crafted radiomic descriptors and patient-level clinical metadata in the form of data-attention. This allows for classification (HGG versus LGG) and severity scoring to be performed concurrently. For decisor transparency, both spatial and channel-level explanations are included using Grad CAM++ and SHAP. The model is trained and tested on BraTS 2021 dataset with 98.34% accuracy, 97.73% F1-score and MAE=0.38 for severity prediction. This paper provides novel insights into the clinical interpretability of multimodal fusion and attention-based weighting, in addition to its effect on the predictive performance. Ablation study and comparisons with state-of-the-arts demonstrate the necessity and effectiveness of each component. The incorporation of explainableAI techniques builds trust and improves usability in clinical workflows, making NeuroExplainAI an appealing platform for reliable, interpretable, and individualized brain tumor assessment.
Journal of Theoretical and Applied Information Technology (Mon,) studied this question.