Authors: Dinnesh Gr, Manoj Jai Sudhan, Mrs. A. Jeyanthi, Mrs.G.Priyaa Sri
Abstract: Stroke is a major global health challenge, contributing significantly to mortality and disability, and placing a heavy burden on healthcare systems. Timely and accurate diagnosis is critical to mitigate long-term complications and improve patient outcomes. This study introduces a hybrid deep learning framework for automated stroke detection in brain CT images, integrating Vision Transformer (ViT), LASSO regression, and DenseNet121 to enhance diagnostic accuracy and efficiency. Utilizing a Kaggle dataset of 1900 CT images (950 stroke, 950 normal), the system employs preprocessing techniques, including resizing to 224×224 pixels, grayscale-to-RGB conversion, and data augmentation (flipping, rotation, blurring), to ensure model robustness and adaptability. The ViT model extracts high-level semantic features, capturing global dependencies through self-attention mechanisms, which are then refined using LASSO regression for feature selection to reduce dimensionality and prevent overfitting. The refined features are fed into DenseNet121, a convolutional neural network optimized for efficient parameter usage and gradient flow, for binary classification (stroke vs. normal). A Tkinter-based graphical user interface facilitates seamless interaction, allowing radiologists to upload images and receive real-time predictions, enhancing clinical workflows. The system is designed for scalability, local deployment, and integration with hospital systems like PACS, addressing challenges of diagnostic delays and inter-observer variability. Evaluation on the dataset demonstrates robust performance, with an accuracy of 92.69%, precision of 91.36%, recall of 94.03%, and F1-score of 92.68%. These metrics underscore the system’s reliability in minimizing false negatives, critical for clinical applications. This framework advances automated stroke diagnosis by combining transformer and convolutional architectures, offering a scalable, interpretable solution for emergency settings and laying the groundwork for future enhancements in multi-class stroke classification and real-time deployment.
