Early diagnosis of ischemic stroke on non-contrasted CT scan using machine learning. A case study of Mulago National Referral Hospital

Abstract

Background Ischemic stroke is a major cause of disability and death, especially in low-resource settings such as Uganda. Timely diagnosis using non-contrast CT scans is essential but often delayed due to the limited number of radiologists and the complexity of image interpretation. This study developed and tested a machine learning model to support faster and more accurate detection of ischemic stroke at Mulago National Referral Hospital. Methods A dataset of 1,000 non-contrast CT scans was collected and annotated by three radiologists. The images were preprocessed through intensity normalization and texture feature extraction. A convolutional neural network (CNN) was trained and validated using five-fold cross-validation. Model performance was measured using sensitivity, specificity, accuracy, and area under the receiver operating characteristic curve (AUC). A web-based platform and RESTful API were also developed to enable clinical testing and deployment. Results The CNN achieved an average accuracy of 99.4%, with 100% sensitivity and 98.9% specificity across all validation folds. The AUC was 0.995, showing excellent discrimination between stroke and non-stroke images. When compared to a senior radiologist, the model detected all stroke cases correctly, producing results within three seconds per scan. The web interface allowed image upload, automatic analysis, and generation of a diagnostic summary report. Conclusion The developed model accurately detects ischemic stroke on CT scans and performs comparably to an experienced radiologist while offering faster results. The accompanying platform demonstrates readiness for integration into hospital workflows. This study highlights the potential of machine learning to improve diagnostic efficiency and stroke care in Uganda and similar healthcare settings.