Combining Non-Contrast CT Radiomics with MR PREDICTS Improves Prediction of Futile Recanalization after Mechanical Thrombectomy.

PURPOSE: Despite complete recanalization (eTICI 3) following mechanical thrombectomy (MT), many patients with acute ischemic stroke (AIS) due to large vessel occlusion (LVO) experience futile recanalization (FR). We evaluated whether baseline non-contrast CT (NCCT) radiomics, alone or combined with MR PREDICTS variables, improves prediction of post-MT functional and imaging outcome.

MATERIALS AND

METHODS

We retrospectively analyzed 167 consecutive patients with anterior-circulation AIS-LVO who achieved eTICI 3 reperfusion after MT and underwent baseline NCCT, multiphase CTA and 24-hour NCCT. Two ASPECTS-level slices were selected from baseline NCCT for each patient, and 377 radiomic features were extracted from each slice. Nine classifiers were trained and ensemble machine learning (EML) models were generated by averaging the predicted probabilities of classifiers with test-set accuracy >65%. The primary endpoint was FR, defined as 90-day modified Rankin Scale score >2. The secondary endpoint was ischemic changes on 24-hour NCCT. Data were split into training and test set in a 70:30 ratio.

RESULTS

For FR, the radiomics-only EML showed poor discrimination (AUC 0.54, 95% CI 0.38-0.71). Adding MR PREDICTS improved AUC to 0.69 (95% CI 0.55-0.84) with 0.71 accuracy, 0.77 sensitivity, 0.64 specificity, but without significant difference (DeLong p = 0.10). For 24-hour NCCT ischemic changes, the radiomics-only and combined models achieved an AUC of 0.73 (95% CI, 0.58-0.88) and 0.78 (95% CI, 0.65-0.92) respectively, with no significant difference (DeLong p = 0.37).

CONCLUSIONS

Baseline NCCT radiomics reliably predicted early imaging evolution after MT, whereas FR prediction improved only after integration with MR PREDICTS. Prospective multicenter validation is needed.