Robust slice-level stroke classification in non-contrast head CT via structural consistency regularization and counterfactual suppression.

Background: Stroke is a time-critical neurological condition with abrupt onset, rapid progression, and a narrow treatment window. In emergency settings, non-contrast head CT is widely used for initial screening and preliminary subtyping, but single-slice CT interpretation remains challenging because lesions often show low contrast, blurred boundaries, diverse morphologies, and interference from skull-related structures and imaging artifacts. Methods: We propose a slice-level stroke CT classification framework that integrates structural consistency regularization, counterfactual suppression, and multi-branch feature refinement. The consistency branch enforces representation stability under perturbations and structure-preserving transformations, while the counterfactual branch attenuates spurious cues and refines branch-wise evidence before confidence-aware aggregation. Results: Comparative experiments on two stroke CT slice datasets show that the proposed method achieves the best overall performance in terms of Accuracy, Precision, Recall, F1, and AUC. On the primary dataset, it reaches Acc = 0.9823, Precision = 0.9766, Recall = 0.9740, F1 = 0.9805, and AUC = 0.9982. On the secondary dataset, it reaches Acc = 0.9652, Precision = 0.9688, Recall = 0.9773, F1 = 0.9698, and AUC = 0.9989. Conclusion: These results indicate that the proposed method can more reliably identify stroke-discriminative evidence under low-contrast and high-interference conditions in the single-slice setting, thereby providing more robust support for emergency stroke triage, preliminary diagnosis, and lesion-aware imaging feedback for intelligent neurorehabilitation applications.