Minimizing Human-Induced Variability in Quantitative Angiography for Robust and Explainable AI-Based Occlusion Prediction

Reducing human error improves AI reliability: The study found that variability in how contrast dye is injected during angiograms creates bias that undermines AI predictions. By standardizing this injection process mathematically, the researchers improved the consistency and trustworthiness of AI-based medical predictions for aneurysm treatment outcomes.

Technical debiasing is necessary for clinical AI: The team used a mathematical technique (deconvolution and reconvolution) to remove injection-related bias from imaging data before training their AI model on 458 patient cases. This demonstrates that responsible AI in healthcare requires identifying and eliminating sources of human-introduced variability, not just building better algorithms.

Explainability is essential for clinical adoption: The researchers used LIME (Local Interpretable Model Agnostic Explanations) to identify which specific imaging features the AI model relied on when predicting treatment success. This transparency helps clinicians understand why the AI made a prediction, making it more trustworthy and suitable for real-world medical decision-making.

AI can predict treatment outcomes reliably: When properly debiased, the deep neural network successfully predicted six-month occlusion status (healing of brain aneurysms treated with flow diverters), suggesting that responsible AI development can deliver practical clinical value for patient care planning.