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

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Medicine
• Cerebrovascular Surgery

Background:

• Contrast injection variability in quantitative angiography (QA) poses a challenge for accurate intracranial aneurysm (IA) occlusion prediction using deep neural networks (DNNs).
• Addressing this bias is crucial for reliable AI-driven outcome prediction in IA treatment.

Purpose of the Study:

• To implement an algorithm for removing injection bias in QA data to reduce variability.
• To assess the impact of explainable AI (XAI) on the interpretability and reliability of DNN models for predicting IA occlusion after flow diverter treatment.

Main Methods:

• Angiograms from 458 patients with flow diverter-treated IAs were analyzed with 6-month occlusion status as the outcome.
• Injection variability was minimized by deconvolving and reconvolving the parent artery input with a standardized curve.
• A DNN predicted 6-month occlusion, with Local Interpretable Model-Agnostic Explanations (LIME) used for feature identification.

Main Results:

• The DNN using uncorrected QA achieved an AUROC of 0.60 and accuracy of 0.58.
• After injection bias correction, the DNN's AUROC increased to 0.79 and accuracy to 0.73.
• LIME plots were incorporated to enhance the interpretability of model predictions.

Conclusions:

• Standardizing QA parameters through injection bias correction enhances the accuracy of IA occlusion prediction.
• Integrating XAI, such as LIME, improves the transparency and clinical relevance of AI models for outcome prediction.