Machine Learning-Enabled Automated Determination of Acute Ischemic Core From Computed Tomography Angiography

Area of Science:

• Neurology
• Radiology
• Artificial Intelligence

Background:

• Advanced neuroimaging interpretation for endovascular stroke therapy (EST) is limited by availability and expertise.
• Computed tomography perfusion (CTP) is a key advanced imaging modality for stroke evaluation.
• Computed tomography angiography (CTA) is more widely accessible but has limitations in detailed analysis.

Purpose of the Study:

• To develop and validate an automated machine learning (ML) method to detect large vessel occlusion (LVO) and ischemic core volume using CTA.
• To assess the accuracy of ML-based CTA analysis compared to CTP-RAPID definitions.
• To determine if CTA alone contains sufficient information for accurate neuroimaging evaluation for EST.

Main Methods:

• A novel convolutional neural network, DeepSymNet, was created and trained on CTA source images.
• The model was trained to identify LVO and infarct core, validated against CTP-RAPID definitions.
• Performance was evaluated using 10-fold cross-validation and receiver-operative curve area under the curve (AUC) statistics on 297 patients.

Main Results:

• DeepSymNet achieved an AUC of 0.88 for LVO detection.
• The algorithm determined infarct core with AUCs of 0.88 and 0.90 (for ≤30 mL and ≤50 mL thresholds, respectively).
• High accuracy was maintained across early (0-6 hours) and late (6-24 hours) time windows, with AUCs up to 0.91.

Conclusions:

• Machine learning analysis of CTA can accurately identify LVO and estimate ischemic core volume.
• This automated ML approach may provide comparable accuracy to advanced imaging modalities for EST evaluation.
• CTA, when analyzed by ML, holds significant potential for streamlining the neuroimaging workup for stroke patients eligible for endovascular therapy.