Interpretable Aneurysm Classification via 3D Concept Bottleneck Models: Integrating Morphological and Hemodynamic Clinical Features

Concept Bottleneck Models offer a more trustworthy alternative to traditional AI "black boxes" in medical imaging: Rather than using post-hoc explanation methods like saliency maps that show correlations after the fact, this framework constrains the AI's decision-making to align with established clinical concepts that neurosurgeons already understand, addressing a key barrier to clinical adoption and regulatory approval.

The model achieves high diagnostic accuracy (93.33% ± 4.5%) while maintaining transparency: By mapping complex 3D brain imaging data to interpretable morphological and hemodynamic features, the system demonstrates that explainability doesn't require sacrificing predictive performance—a critical finding for responsible medical AI deployment.

The framework integrates multiple types of clinical information into a unified decision process: By combining both morphological features (physical structure) and hemodynamic features (blood flow characteristics) through a structured "bottleneck" layer, the model mirrors how clinicians actually think about aneurysm assessment, making its reasoning more aligned with established neurosurgical principles.

This approach addresses a fundamental regulatory and ethical challenge in medical AI: The paper directly tackles why many accurate AI models haven't been adopted clinically—the lack of transparency—by building interpretability into the model's core architecture rather than trying to explain it after the fact, which is essential for gaining clinician trust and regulatory approval.