Abstract
Stroke is the second leading cause of death globally. Computed Tomography plays a significant role in the initial diagnosis of suspected stroke patients. Currently, stroke is subjectively interpreted on CT scans by domain experts, and significant inter- and intra-observer variation has been documented. Several methods have been proposed to detect ischemic brain stroke automatically on CT scans using machine learning and deep learning, but they are not robust and their performance is not ready for clinical practice. We propose a fully automatic method for acute ischemic stroke detection on brain CT scans. The system's first component is a brain slice classification module that eliminates the CT scan's upper and lower slices, which do not usually include brain tissue. In turn, a brain tissue segmentation module segments brain tissue from CT slices, followed by tissue contrast enhancement using the Extreme-Level Eliminating Histogram Equalization technique. Finally, the processed brain tissue is classified as either normal or ischemic stroke using a classification module, to determine whether the patient is suffering from an ischemic stroke. We leveraged the use of the pre-trained ResNet50 model for slice classification and tissue segmentation, while we propose an efficient lightweight multi-scale CNN model (5S-CNN), which outperformed state-of-the-art models for brain tissue classification. Evaluation included the use of more than 130 patient brain CT scans curated from King Fahad Medical City (KFMC). The proposed method, using 5-fold cross-validation to validate generalization and susceptibility to overfitting, achieved accuracies of 99.21% in brain slice classification, 99.70% in brain tissue segmentation, 87.20% in patient-wise brain tissue classification, and 90.51% in slice-wise brain tissue classification. The system can assist both expert and non-expert radiologists in the early identification of ischemic stroke on brain CT scans.