Predicting Acute Ischemic Stroke Core in Multiphase CT Angiography Using a CT Perfusion-Trained Neural Network

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neurology

Background:

• Accurate identification of the ischemic stroke core is critical for timely treatment decisions.
• CT angiography (CTA) is accessible but lacks standardized stroke core prediction methods.
• CT perfusion (CTP) provides better tissue classification but is costly and complex.

Purpose of the Study:

• To develop a convolutional neural network (CNN) capable of identifying ischemic stroke core in multiphase CTA (mCTA) images.
• To utilize CTP-derived core maps as ground truth for training the CNN.
• To assess the model's performance across varying stroke severities using derived datasets.

Main Methods:

• Retrospective analysis of 99 acute ischemic stroke patients with both CTP and mCTA.
• CNN training using CTP core maps and application to mCTA feature maps.
• Performance evaluation using Dice Similarity Coefficient (DSC), accuracy metrics, and precision-recall curves.

Main Results:

• The CNN achieved a DSC of 0.43 on the base dataset for distinguishing stroke core from healthy tissue.
• Performance improved with increasing stroke severity, with the Severe20 dataset yielding a DSC of 0.60.
• Thresholding analysis and precision-recall curves provided insights into model performance trade-offs.

Conclusions:

• Machine learning can integrate CTP's core detection with CTA's accessibility for improved stroke imaging.
• The developed model shows potential for enhancing stroke diagnosis, particularly in severe stroke cases.
• Further optimization and validation on larger datasets are needed for clinical translation.