Stroke is a serious neurological condition that causes high morbidity and mortality rates worldwide. Rapid and accurate diagnosis is critical to improving patient treatment and quality of life. To address this issue, two lightweight deep learning models based on MobileNetV2 were designed for stroke diagnosis and classification. The first model combines channel and spatial attention mechanisms by integrating CBAM (Convolutional Block Attention Module) into MobileNetV2. The second model applies a multi-attention strategy using ECA (Efficient Channel Attention) and Coordinate Attention blocks. Transfer learning was applied using a pre-trained MobileNetV2 network on ImageNet for both models, preventing overfitting on small medical datasets. The results demonstrate that attention mechanisms are effective in capturing subtle differences between classes. Model 1 (99.4% accuracy) and Model 2 (99.8% accuracy) demonstrated high performance, supported by precision, recall, F1-score, and AUC values. The primary goal in selecting MobileNetV2 was to ensure that the models could be integrated into devices with limited resources, such as smartphones and embedded systems. The results were found to be promising in terms of both high accuracy and low computational cost. The proposed approach has the potential to provide scalable and real-time solutions, making a significant contribution to stroke detection.
Sibel Barın Özkan (Tue,) studied this question.