Interpretable Deep Learning-Based AI Framework with Multi-Sequence Attention for Brain Tumor Subtype Classification in MRI Scans

Robust and interpretable classification of brain tumor subtypes remains a core challenge in medical image analysis due to tumor heterogeneity, modality-specific contrast variations, and limited generalizability of conventional models. This research proposes a deep learning-based framework leveraging Convolutional Neural Networks (CNN), Transformer-based architectures, and multi-sequence attention mechanisms to classify glioma subtypes using T1, T1c, T2, and FLAIR MRI sequences from the BraTS 2023 dataset. The attention-enhanced CNN model achieved state-of-the-art performance with an F1-score of 0.974, AUC of 0.982, and classification accuracy of 98.2%, outperforming baseline CNNs (F1 = 0.88) and radiomics-based SVM models (F1 = 0.84). Integration of spatial and inter-sequence attention enabled dynamic weighting of modality-specific features, enhancing diagnostic precision and model interpretability. Grad-CAM and SHAP-based attribution maps showed a 90% overlap with expert-defined tumor regions, and a mean interpretability rating of 4.8/5 from clinical reviewers. The system maintained <150ms inference latency, meeting real-time diagnostic constraints. Transformer models demonstrated marginally higher accuracy but required 40–60% more compute resources, limiting their deployment feasibility. Early stopping at epoch 25 effectively minimized overfitting and preserved generalization across institutional data sources. Comparative benchmarking against ARIMA, SVM, and prior CNN architectures validated the superiority of attention-integrated deep networks in high-dimensional, multi-modal neuroimaging classification. The findings establish a scalable and clinically deployable AI framework for brain tumor subtype discrimination, with future directions including hybrid radiomics-transformer models and federated learning for decentralized clinical deployment.