StrokeNet-CT: An EfficientNetV2-Based Framework with Adaptive Windowing and Attention Mechanisms for Binary Stroke Classification on NCCT Images

Stroke is still one of the most common neurological emergencies, and rapid image interpretation plays a major role in determining the treatment a patient can receive. Although Non-Contrast CT (NCCT) is the first imaging choice in most emergency rooms, the early signs of ischemia are often so subtle that they are easily overlooked. In this work, we introduce StrokeNet-CT is a model based on EfficientNetV2B0 that has been modified to better suit NCCT characteristics. Two additional components are the Dynamic Windowing Module which adjusts image contrast automatically and the Stroke- Focused Attention Block which highlights areas where faint ischemic patterns may occur were incorporated to address well- known limitations of NCCT. The model was trained on 4,829 images and demonstrated an accuracy of 91.40%, outperforming standard CNN architectures trained under the same conditions. While the model is not intended to replace radiologists,the results suggest that it could serve as a practical aid in settings where specialized imaging or experienced readers are not always available.