Cerebrovascular CTA radiomics for objective collateral grading in acute ischemic stroke.

Segmentation models accurately annotated cerebral arteries and circle of Willis segments.

• Arterial tree segmentation achieved a 95th percentile Hausdorff distance of 4.49 and Dice similarity coefficient of 0.83.
• Circle of Willis multiclass segmentation achieved a 95th percentile Hausdorff distance of 2.27 and Dice similarity coefficient of 0.81.
• nnU-Net models were trained on 40 arterial tree CTAs and 125 multiclass circle of Willis cases.

The vessel-tree radiomics model outperformed the atlas-based MCA mask model on the internal test set.

• Vessel-tree model achieved an AUROC of 0.88 on the internal test set.
• MCA mask model achieved an AUROC of 0.82 on the internal test set.
• The internal test set consisted of 69 patients from the MR CLEAN trial.
• After feature selection, 6 top features were identified for the vessel-tree radiomics model compared to 98 for the MCA mask-based model.

The vessel-tree radiomics model substantially outperformed the MCA mask model on the external validation cohort.

• Vessel-tree model achieved an AUROC of 0.83 on the external test set.
• MCA mask model achieved an AUROC of 0.66 on the external test set.
• The external cohort consisted of 140 acute LVO patients.
• The difference in performance (0.83 vs 0.66) demonstrates superior generalizability of the vessel-tree approach.

Adding circle of Willis features to the vessel-tree model further improved collateral score prediction performance.

• The combined vessel-tree/CoW model achieved an AUROC of 0.87.
• 32 features were identified after selection for the combined vessel-tree/CoW model.
• The combined model demonstrated improved accuracy compared to the vessel-tree model alone on external validation.

The study used a retrospective cohort of 343 LVO patients from the MR CLEAN trial split into training/validation and testing sets.

• The training/validation set comprised 274 patients and the testing set comprised 69 patients.
• Admission CTAs were analyzed retrospectively.
• The classifier was trained to distinguish sufficient (>50%) from insufficient (≤50%) collateral status according to the Tan score system.
• A random forest classifier was used for final collateral status classification.

Collateral grading in acute ischemic stroke is rater-dependent and error-prone, motivating the development of automated assessment tools.

• Collateral circulation is described as 'a key determinant of functional outcome after large vessel occlusion (LVO) and informs thrombectomy decisions.'
• Current grading is described as 'rater-dependent and error-prone.'
• Automated scoring has potential to 'reduce inter-rater variability, improve workflow efficiency, and support individualized treatment decisions.'