Vision Transformer Model for Efficient Stroke Detection in Neuroimaging

A brain stroke occurs when blood flow to a part of the brain is disrupted, potentially caused by a blocked or ruptured blood vessel. Deprived of oxygen and nutrients, brain cells can start dying within minutes, leading to irreversible damage. Early diagnosis and treatment are crucial to minimize brain damage and improve recovery chances. Clinical assessments and imaging techniques like Computed Tomography (CT) and Magnetic Resonance Imaging (MRI) scans are commonly used for rapid detection, but manual analysis has limitations, including delays and subjectivity. AI-based models offer a faster and more consistent approach for stroke diagnosis, enhancing accuracy. In this study, an explainable Vision Transformer (ViT) model is proposed for stroke classification and localization from brain CT images. The model is validated on a dataset of 6,651 samples. To address an unbalanced dataset, two training scenarios were employed. Scenario-1 directly used the unbalanced dataset, while Scenario-2 equalized sample numbers through data augmentation. In the test phase, Scenario-1 achieved 97.25% accuracy, 98.46% precision, 96.00% recall, 98.50% specificity, and a 97.22% F-1 score. In contrast, Scenario-2 achieved even higher performance with 98.75% accuracy, 99.49% precision, 98.00% recall, 99.50% specificity, and a 98.74% F-1 score. Analyzing the softmax ratios in the model predictions revealed that Scenario-2, with synthetic images in the training set, produced more reliable results. The study also used the Grad-CAM algorithm to visualize the areas of focus in the models' predictions, showcasing their superior localization capabilities. This proposed model is well-suited for clinical use due to its high accuracy rates and robust localization abilities, potentially improving stroke diagnosis and treatment outcomes.