Stroke Classification in Brain Computed Tomography Images Using Vision Transformers and GAN-based Data Augmentation

This study presents an innovative approach to stroke classification. The research utilizes brain computed tomography (CT) images to distinguish between three classes: “no stroke” “ischemic stroke” and “hemorrhagic stroke” employing Vision Transformers (ViTs), a deep learning-based method incorporating attention mechanisms. In this work, ViTs were effectively applied as a powerful method for image-based classification. To enhance model performance, various training strategies and data augmentation techniques were implemented. Specifically, GAN-based architectures such as SRGAN (Super-Resolution GAN) and BSRGAN (Blind Super-Resolution GAN) were used to expand the dataset and improve its diversity. These GAN-based augmentation techniques significantly improved the model’s overall performance and classification accuracy. The Vision Transformer model was rigorously evaluated through multi-class classification tasks using a range of performance metrics. In the three-class classification task, the model achieved 99.06% accuracy, 98.18% precision, 98.94% recall, and a 98.54% F1-score. For the binary classification of ischemic vs. hemorrhagic stroke, the model reported 99.78% accuracy, 99.02% precision, 99.66% recall, and a 99.26% F1-score. In the binary classification of stroke presence, the model achieved 98.68% accuracy, 97.80% precision, 98.54% recall, and a 98.14% F1-score. These findings demonstrate the potential of Vision Transformers to assist in faster and more reliable stroke diagnosis and highlight their contribution to the development of decision support systems in medical applications.