Detecting hemorrhagic stroke from computed tomographic scans using machine learning models comparison

Abstract

Introduction: Stroke is the most leading cause of death and disability worldwide, with hemorrhagic stroke being the most dangerous due to bleeding in the brain. To minimize the impacts, early detection is crucial for effective management and timely intervention. This is precisely the motivation behind our research, which aims to develop a reliable and rapid diagnostic support system. Methods: In this study, the authors combined machine learning (ML) models to detect stroke using a dataset of computerized tomography (CT) images. The study was conducted on a real database containing CT images collected from Moroccan patients. The method used in data organization and preprocessing were performed, followed by feature extraction from each image, such as intensity, grayscale, and histogram characteristics. These extracted features were then compressed using several algorithms, including Principal Component Analysis (PCA). The processed data were fed into the most robust machine learning classifiers based on existing literature. Results: As a result, the XGBoost model achieved the highest classification accuracy, with 93% precision, using a Leave-One-Subject-Out (LOSO) validation scheme. Conclusion: This study is a step forward in improving patient healthcare by enabling early detection, which could lead to timely, potentially life-saving interventions.