Minimizing human-induced variability in quantitative angiography for a robust and explainable AI-based occlusion...

Bias from contrast injection variability is a significant obstacle to accurate intracranial aneurysm (IA) occlusion prediction using quantitative angiography (QA) and deep neural networks (DNNs). This study explores bias removal and explainable AI (XAI) for outcome prediction.

To implement an injection bias removal algorithm for reducing QA variability and examine the impact of XAI on the reliability and interpretability of deep learning models for occlusion prediction in flow diverter-treated aneurysms.

This study used angiograms from 458 patients with flow diverter-treated IAs, with 6-month follow-up defining occlusion status. We minimized injection variability by deconvolving the parent artery input to isolate the aneurysm's impulse response, then reconvolving it with a standardized injection curve. A DNN trained on these QA-derived biomarkers predicted 6-month occlusion. Local Interpretable Model-Agnostic Explanations (LIME) identified the key imaging features influencing the model, ensuring transparency and clinical relevance.

The DNN trained with uncorrected QA parameters achieved a mean area under the receiver operating characteristic curve (AUROC) of 0.60±0.05 and an accuracy of 0.58±0.03. After correcting for injection bias by deconvolving the parent artery input and reconvolving it with a standardized injection curve, the DNN's AUCROC increased to 0.79±0.02 and accuracy to 0.73±0.01. Sensitivity and specificity were 67.61±1.93% and 76.19±1.12%, respectively. LIME plots were added for each prediction to enhance interpretability.

Standardizing QA parameters via injection bias correction improves occlusion prediction accuracy for flow diverter-treated IAs. Adding explainable AI (eg, LIME) clarifies model decisions, demonstrating the feasibility of clinically interpretable AI-based outcome prediction.