Deep learning and machine learning for differentiation between contrast extravasation and hemorrhagic transformation in post-thrombectomy stroke CT.

BACKGROUND

Mechanical thrombectomy (MT) improves outcomes in acute ischemic stroke (AIS) but often results in hyperdensities on non-contrast CT (NCCT), which may represent either hemorrhagic transformation (HT) or contrast extravasation (CE). Distinguishing between them is critical, as HT may contraindicate early anticoagulation. We developed and validated machine learning (ML) models to differentiate HT from CE using NCCT within 6 h post-MT.

METHODS

We retrospectively analyzed 351 patients with anterior circulation AIS who underwent MT. Among 111 patients with post-MT hyperdensities, follow-up CT (24-72 h) classified them as HT (n = 41), CE (n = 34), or mixed HT+CE (n = 36). Radiomics-based models-Support Vector Machine (SVM), Random Forest (RF), and Logistic Regression (LR)-were trained on 72 patients using segmented hyperdensities. A U-Net deep learning model was trained on raw axial slices. Performance was assessed on a 39-patient test set using accuracy, sensitivity, specificity, F1-score, and AUC.

RESULTS

The mean age was 65 years, and baseline characteristics were largely comparable across groups: 71% were male; 50% had a history of hypertension; and mean ASPECTS score was 6.7. All models demonstrated high classification performance, with U-Net achieving the highest overall accuracy (96%) and F1-score (0.96). Sensitivity/specificity for HT were: SVM (94.1%/97.0%), LR (94.1%/100.0%), RF (82.4%/91.2%), and U-Net (82.4%/94.1%). For CE: SVM (92.3%/97.3%), LR (100.0%/97.3%), RF (91.7%/94.1%), U-Net (100.0%/100.0%). For HT+CE: SVM (70.0%/91.2%), LR (90.0%/97.1%), RF (70.0%/88.2%), U-Net (100.0%/100.0%). U-Net significantly outperformed RF across all metrics (p < 0.01), but was not significantly different from SVM or LR.

CONCLUSION

ML models applied to NCCT can accurately differentiate CE from HT post-MT. While U-Net offers advantages in learning from raw imaging data, traditional models performed comparably. Larger studies and hybrid approaches integrating radiomics and deep learning are warranted.