Factors influencing the reliability of a CT angiography-based deep learning method for infarct volume estimation

Area of Science:

• Neuroimaging
• Machine Learning in Medicine
• Stroke Imaging

Background:

• CT angiography (CTA)-based machine learning methods often overestimate infarct core and final infarct volumes (FIV).
• Understanding factors influencing the reliability of these estimations is crucial for accurate stroke assessment.

Purpose of the Study:

• To assess factors influencing the reliability of CTA-based machine learning methods for infarct volume estimation.
• To evaluate the impact of collateral circulation and CTA timing on the accuracy of convolutional neural network (CNN) estimations.

Main Methods:

• Assessed the correlation between CNN-estimated volumes and FIV using Pearson correlation coefficients in 121 acute ischemic stroke patients.
• Evaluated the effect of collateral status (Miteff system, HIR) and thrombectomy outcomes (successful vs. futile) on estimation accuracy.
• Analyzed the timing of CTA acquisition relative to CTP studies.

Main Results:

• Correlation between CNN estimations and FIV was poor to moderate (r=0.09-0.50) and not significantly affected by collateral status.
• The strongest correlation was observed in patients with futile thrombectomies (r=0.61).
• CNN estimates showed a trend towards overestimation compared to FIVs, with CTA acquired in the mid-arterial phase for most patients.

Conclusions:

• Collateral status does not significantly affect the reliability of CNN-based infarct volume estimation.
• CTA timing in the mid-arterial phase appears to be a key factor contributing to infarct volume overestimation.
• Optimization of CTA protocols is necessary for improving the accuracy of machine learning-based infarct core estimation.