Stroke detection and classification from computed tomography (CT) remains a critical and challenging task in medical imagingdue to the complexity of lesion patterns, noise variations and unbalanced datasets. In this study, we propose a novel hybriddeep learning model, StrokeFuse-AttnNet, which integrates both global (ResNet50) and local (DenseNet121) convolutionalfeature extractors with a self-attention mechanism to improve spatial focus and semantic interpretability. A hierarchical featurefusion strategy concatenates multi-scale features, which are then processed by a self-attention module to highlight key strokeregions and reduce irrelevant activations. We use data augmentation and SMOTE on training samples to address imbalanceand improve generalization. The proposed model was evaluated on both publicly and privately available brain CT datasets.StrokeFuse-AttnNet achieved an accuracy of 98.27% and an AUC of 0.983 on the public dataset and an accuracy of 96.04%and an AUC of 0.9501 on the private dataset. The results show that the model has higher accuracy, reliability and generalizationthan existing and baseline methods. The proposed model is lightweight, with only 32 million parameters and can be used inreal-time clinical diagnostic processing systems that require 40 GFLOPs. The model has the potential to support radiologistsin the efficient and rapid diagnosis of strokes on non-contrast CT images.