Machine and Deep Learning Approaches for Accurate Identification of Ischemic and Hemorrhagic Strokes

Timely and accurate detection of ischemic and hemorrhagic strokes remains one of the most critical challenges in neuroimaging. This study proposes a systematic deep-learning framework for multi-class medical-image classification, concentrating on three popular convolutional neural network structures: a basic CNN, VGG-16, and VGG-19. All models pass through a common training and evaluation pipeline that adopts categorical cross-entropy loss for multi-class outcomes and the adam optimizer to update weights in a data-efficient manner. Performance is measured using accuracy, precision, recall, and f1-score, with separate trials on MRI scans, CT images, and merged MRI-CT sets. The baseline CNN establishes an initial performance level, against which gains from the deeper networks can be assessed. VGG-16 and VGG-19, characterized by smaller receptive fields and a more intricate layer organization, consistently exceed the simpler variant on every metric. VGG-16 strikes a practical balance between model depth and computational load, whereas VGG-19, by adding three extra convolutional layers, provides a modest lift in classification accuracy. In the analysis of unseen test images, the softmax layer generates a normalized probability distribution, and the argmax function then selects the class with the highest score as the final prediction. Through both training and evaluation, a precise mathematical framework describes the construction of the loss, the subsequent gradients required for optimizer updates, and the metrics employed to measure classification accuracy. Such documentation offers researchers a clear and replicable map of how the network interprets input and formulates its output. Empirical tests reveal that both VGG-16 and VGG-19 surpass classic CNN architectures in extracting subtle features from medical scans. The two, VGG-19 consistently yields the top accuracy and stability across several validation sets. Taken together, these findings endorse the use of deeper VGG networks as dependable tools for advancing clinical image analysis and diagnostics.