The combined intrathrombus-perithrombus CT-based radiomics model with logistic regression effectively predicts complete recanalization after endovascular thrombectomy, achieving AUC values of 0.93 (training), 0.88 (testing), and 0.86 (validation), outperforming single-region models.
Key Findings
Results
The combined intrathrombus-perithrombus radiomics model with logistic regression outperformed single-region models in predicting complete recanalization after EVT.
AUC values were 0.93 in the training cohort, 0.88 in the internal testing cohort, and 0.86 in the external validation cohort
Complete recanalization was defined as mTICI 2c/3 following endovascular thrombectomy
Eleven different machine learning classifiers were tested, with logistic regression achieving the best performance
The combined model incorporated features from both non-contrast CT and CT angiography (CTA)
Results
The perithrombus region contributed the majority of features to the combined radiomics model.
10 of 15 total features in the combined model were derived from the perithrombus region
Only 5 of 15 features originated from the intrathrombus region
A total of 428 radiomics features were initially extracted before feature selection
LASSO (Least Absolute Shrinkage and Selection Operator) regression was used for feature selection
Results
Decision curve analysis confirmed superior clinical utility of the combined model over single-region models.
Decision curve analysis was used to assess net clinical benefit across threshold probabilities
The combined intrathrombus-perithrombus model demonstrated greater net benefit compared to single-region models
Clinical utility assessment supported the combined model as the preferred approach for patient selection and treatment optimization
Methods
The study included 406 AIS patients with anterior circulation large-vessel occlusion from three centers spanning December 2018 to April 2024.
Patients were allocated to training (n = 178), internal testing (n = 77), and external validation (n = 151) cohorts
The study was retrospective in design
Three centers contributed to the dataset, enabling external validation
The study population was restricted to anterior circulation large-vessel occlusion cases
Methods
Radiomics features were extracted from two CT imaging modalities covering two distinct anatomical regions around the thrombus.
Features were extracted from both non-contrast CT and CT angiography (CTA)
Two regions of interest were defined: intrathrombus and perithrombus
A total of 428 radiomics features were extracted in total before selection
LASSO regression reduced the feature set to 15 features for the final combined model
What This Means
This research suggests that artificial intelligence analysis of CT scan images taken before stroke treatment can accurately predict whether a blood clot removal procedure (endovascular thrombectomy) will fully restore blood flow to the brain. The researchers analyzed patterns in two areas of CT images — within the blood clot itself (intrathrombus) and in the tissue immediately surrounding the clot (perithrombus) — in 406 stroke patients from three different hospitals. By combining information from both regions, their model was able to predict complete blood flow restoration with high accuracy (AUC of 0.86–0.93 depending on the patient group tested).
A notable finding was that the area surrounding the clot, rather than the clot itself, provided the most informative imaging features — 10 out of the 15 features used in the final model came from the perithrombus region. Among eleven different machine learning approaches tested, logistic regression performed best. The model also demonstrated real-world clinical value through decision curve analysis, meaning it could genuinely help doctors make better treatment decisions rather than just performing well on statistical benchmarks.
This research suggests that routinely acquired CT images, which are already standard practice in stroke care, contain rich information that computers can use to help predict treatment outcomes before a procedure begins. If validated in future prospective studies, this type of tool could help clinicians better identify which patients are most likely to benefit from clot removal procedures, potentially improving individualized treatment planning in acute ischemic stroke care.
Zhao Q, Feng S, Jia H, Li M, Tian H, Pan H, et al.. (2026). CT-based intrathrombus and perithrombus radiomics for predicting complete recanalization after endovascular thrombectomy in acute ischemic stroke.. Frontiers in neurology. https://doi.org/10.3389/fneur.2026.1868821