Deep learning and machine learning for differentiation between contrast extravasation and hemorrhagic transformation...

Area of Science:

• Neurology
• Radiology
• Artificial Intelligence

Background:

• Mechanical thrombectomy (MT) is crucial for acute ischemic stroke (AIS) but post-procedure non-contrast CT (NCCT) findings like hyperdensities can be ambiguous.
• Distinguishing hemorrhagic transformation (HT) from contrast extravasation (CE) on NCCT is vital for guiding subsequent anticoagulation therapy.

Purpose of the Study:

• To develop and validate machine learning (ML) models for differentiating HT from CE on NCCT within 6 hours post-MT.
• To assess the performance of radiomics-based (SVM, RF, LR) and deep learning (U-Net) models in this classification task.

Main Methods:

• Retrospective analysis of 351 AIS patients undergoing MT, with 111 showing post-MT hyperdensities classified by follow-up CT.
• Training ML models (SVM, RF, LR, U-Net) on segmented hyperdensities or raw axial slices from a subset of patients.
• Testing model performance on a separate cohort using accuracy, sensitivity, specificity, F1-score, and AUC.

Main Results:

• U-Net achieved the highest accuracy (96%) and F1-score (0.96).
• All models demonstrated high performance; U-Net and Logistic Regression showed excellent sensitivity and specificity for CE (100%).
• U-Net significantly outperformed Random Forest, but performance was comparable to SVM and Logistic Regression.

Conclusions:

• Machine learning models applied to NCCT are effective in distinguishing CE from HT post-MT.
• Deep learning (U-Net) and traditional ML models show comparable diagnostic utility.
• Further research with larger cohorts and hybrid approaches is recommended.