EVM-Fusion: An Explainable Vision Mamba Architecture with Neural Algorithmic Fusion

High accuracy with built-in explainability: EVM-Fusion achieved 99.75% test accuracy on multi-organ medical image classification while incorporating multiple layers of interpretability (spatial attention maps, attention weights, and delta-value visualizations), addressing the critical need for AI systems that doctors can understand and trust in clinical settings.

Multiple pathways provide diverse decision-making insights: The architecture uses three parallel processing pathways (DenseNet, U-Net, and traditional features) that are dynamically combined, allowing clinicians to see how different analytical approaches contribute to the final diagnosis rather than relying on a single "black box" decision path.

Adaptive fusion learns how to combine information: The Neural Algorithmic Fusion mechanism doesn't use fixed rules to combine different analytical pathways; instead, it learns an adaptive algorithm for integration, potentially making the system more flexible across different medical imaging scenarios and organ types.

Designed for real-world medical deployment: By testing on a diverse 9-class multi-organ dataset and emphasizing both accuracy and interpretability, the research addresses practical concerns for responsible AI in healthcare—systems must not only perform well but also provide clinicians with transparent reasoning they can verify and act upon.

Intrinsic explainability over post-hoc explanations: Rather than adding explanations after the fact, EVM-Fusion embeds interpretability throughout its architecture, which is considered more reliable for high-stakes medical applications where understanding why a diagnosis was made is as important as the diagnosis itself.