A Hybrid 1D-CNN and Transformer-Based Deep Learning Framework for Stroke Detection Using Brain CT Scans

Stroke continues to be a major cause of illness and death worldwide, requiring fast and precise diagnosis to enable prompt treatment. Computed Tomography (CT) imaging is widely employed due to its accessibility and speed; however, manual interpretation is time-consuming and requires specialized expertise, which may not be readily available in underserved or high-volume clinical settings. This paper presents a novel hybrid deep learning framework that combines the local pattern recognition of one-dimensional convolutional neural networks (1D-CNN) with the global attention capabilities of Transformer encoders. The proposed architecture employs a multi-branch design comprising a standard convolutional path, a 1D-CNN Transformer fusion block, and a global semantic feature stream. Evaluated on a curated dataset of 437 grayscale brain CT images for binary classification (Normal vs. Stroke), the model achieved 95.88% accuracy, 98.46% sensitivity, 95.1% specificity, 88.9% precision, and an F1-score of 93.4%. The ROC analysis yielded an AUC of 0.991, with a geometric mean of 0.985, reflecting both robustness and generalization capacity. Training converged stably with validation accuracy of 98.3% and low validation loss (0.048), supported by strong regularization strategies. These results surpass conventional CNN baselines, particularly in stroke sensitivity, which is critical for avoiding missed diagnoses. The findings highlight the effectiveness of hybrid architectures in enhancing diagnostic performance and underscore their potential deployment in real-time stroke screening applications, especially in telemedicine and edge-AI environments where rapid triage is essential.