Lightweight Deep Learning for Automated Stroke Detection in Non-Contrast CT Imaging

Stroke remains a major global cause of mortality and long-term disability, underscoring the need for rapid and reliable diagnostic support. Non-Contrast Computed Tomography (NCCT) is the first-line imaging modality in emergency settings, yet early ischemic changes often appear as subtle hypoattenuation, contributing to inter observer variability and delayed diagnosis. This study presents StrokeNet-CT, a lightweight deep learning framework for automated binary stroke detection that distinguishes normal NCCT scans from stroke-affected cases. Built upon the MobileNetV2 architecture, the model leverages transfer learning to mitigate the scarcity of annotated medical data while maintaining a compact footprint suitable for real-time use. The system was trained on 6,650 NCCT images and evaluated on an independent test set, achieving an accuracy of 95.78%, precision of 97.97%, recall of 89.35%, and F1-score of 93.46%. The high precision indicates minimal false positive rates, supporting its use in reducing alert fatigue in clinical workflows. Although the dataset combines ischemic and hemorrhagic strokes without subtype annotation, the proposed framework demonstrates strong potential as an efficient triage-assist tool, particularly in resource-constrained or time-critical environments.