A Deep Transfer Learning Approach for Automated Brain Stroke Detection and Classification

Brain stroke is a serious health problem that causes high death rates and long-term disability. Detecting stroke early using brain CT scans is important but manual checking is slow and can lead to mistakes, especially in emergencies. This study presents an automatic system to detect and classify brain strokes using Convolutional Neural Networks (CNNs). A balanced dataset of 3,360 CT images is used, divided into three groups: Hemorrhagic stroke, Ischemic stroke, and Normal. To improve training, data augmentation methods such as rescaling, rotation, shifting, and zooming are applied. Three CNN models—DenseNet121, InceptionV3, and a custom CNN—are tested and compared. DenseNet121 and InceptionV3 both reached an accuracy of 99.34%, while the custom CNN achieved 99.09%. An ensemble approach combining these models performed the best, achieving an accuracy of 99.39%. Model performance is also measured using confusion matrix, F1-score, precision, recall, and ROC curve. To make the predictions easier to understand, Grad-CAM visualization is applied, which shows the important regions in CT scans that the models focus on. This confirms the clinical usefulness of the system. Compared with previous works, our approach provides higher accuracy and reliability. The results show that deep learning with ensemble and visualization techniques can give strong support to doctors for faster and more accurate stroke diagnosis.