Comprehensive Comparison of Deep Learning Architectures for Stroke Classification From CT Images

Abstract

Stroke, a leading cause of death and permanent disability worldwide, is classified into ischemic and hemorrhagic types. Accurate and timely classification from CT images is critical for effective treatment in emergency care. This study compares modern deep learning models ResNet, ViT, EfficientNet, Inception, ResNeXt, MobileNet, ConvNeXt, ConvNeXtV2, and DaViT-for classifying stroke (ischemic, hemorrhagic) and non-stroke cases from CT images. Models were evaluated using the 2021 Teknofest stroke dataset based on accuracy, precision, specificity, and computational efficiency. Results show that while advanced models like ViT and ConvNeXtV2 offer high performance, lightweight architectures such as MobileNet (F1-score: 97.59%) are clinically viable and ideal for resource-limited environments.