BrainFusionNet: a deep learning and XAI model to understand local, global, and sequential features of MRI images for improved brain tumour detection

Explainability improves clinical trust: The model integrates explainable AI tools (SHAP, LIME, and GradCAM) to visualize which specific regions of MRI images influence tumor detection decisions, addressing a critical need for doctors to understand and verify AI recommendations rather than treating them as "black boxes."

Hybrid architecture outperforms single approaches: BrainFusionNet combines three different deep learning techniques (CNNs, Vision Transformers, and GRUs) to achieve 98% accuracy—outperforming individual state-of-the-art models like DenseNet121 and VGG16 (96% accuracy)—demonstrating that responsible AI development may require combining multiple specialized tools rather than relying on one method.

Better detection of challenging cases: The model shows particular strength in detecting small tumors and complex tumor boundaries that are often obscured by MRI noise, suggesting that thoughtfully designed AI systems can handle real-world medical complexity better than simpler alternatives.

Transparency about data quality matters: The researchers analyzed how MRI image quality affects classification accuracy, emphasizing that responsible AI development requires examining whether input data quality influences model performance—an important consideration for deployment in clinical settings with varying equipment and imaging standards.