Prediction of Major Stroke (NIHSS > 5) Based on MR Perfusion Imaging Using a 3D ResNet

Prediction of stroke symptoms from perfusion patterns in ischemic stroke imaging would be desirable for decision making in relevant clinical scenarios as for example after incomplete mechanical thrombectomy or in perioperative strokes. Aim was to train a model to detect a major stroke defined as NIHSS > 5.

Retrospective study of patients from the prospectively collected local Stroke Registry with large-/medium-vessel anterior circulation or without occlusion from 01/2015-12/2021 (N = 982, median age 74.5 (IQR 64.3-82), 47.5% female). Splits into training-set (80%), validation-set (10%) and temporally separated internal test-set (10%). Admission MR-perfusion Tmax-maps from 6 different MR-scanners were used as input for an 18-layer 3D-ResNet to predict major stroke (NIHSS > 5). Association of reported and predicted NIHSS > 5 with 90-days modified Rankin Scale (mRS) > 2 was tested using multivariate logistic regression.

Vessel occlusion was observed in 80% and NIHSS > 5 in 50% of all patients. On the internal test-set, Receiver Operating Characteristic area under the curve (ROC-AUC) was 0.87 (95%-CI ± 0.04) and Precision-Recall-AUC (PR-AUC) 0.85 (95%-CI ± 0.05). Accuracy was 78%, weighted F1-score 0.78. Including only vessel occlusions in the internal test-set, ROC-AUC was 0.81 (95%-CI ± 0.05) and PR-AUC 0.87 (95%-CI ± 0.05), accuracy was 73%, weighted F1-score 0.73. Reported NIHSS > 5 (aOR 2.03, 95%-CI 0.94-4.38, p = 0.07) and predicted NIHSS > 5 (aOR 3.61, 95%-CI 0.88-14.88, p = 0.08) both predicted 90-days mRS > 2, with similar model performance (BIC: 310.6 vs. 310.7; AIC: 252.4 vs. 252.5).

Prediction of a major stroke (NIHSS > 5) from MR-perfusion Tmax-maps was feasible and performed well on our internal test set.