Attention-enhanced deep learning model for automated classification of hemorrhagic and ischemic stroke on CT imaging.
Authors
Affiliations (9)
Affiliations (9)
- Department of Neurology, Dongzhimen Hospital, Beijing University of Chinese Medicine, Beijing, China.
- Shanxi Bethune Hospital (Shanxi Academy of Medical Sciences), Taiyuan, China.
- Third Hospital of Shanxi Medical University, Taiyuan, China.
- Tongji Shanxi Hospital, Taiyuan, China.
- Beijing University of Technology, Beijing, China.
- Beijing University of Chinese Medicine, Beijing, China.
- Guang'anmen Hospital, China Academy of Chinese Medical Sciences, Beijing, China.
- Department of Neurology, Dongzhimen Hospital, Beijing University of Chinese Medicine, Beijing, China. [email protected].
- Institute for Brain Disorders, Beijing University of Chinese Medicine, Beijing, China. [email protected].
Abstract
Stroke is a disease with extremely high mortality and disability rates worldwide. Hemorrhagic stroke and ischemic stroke require completely different treatment plans, so rapid and accurate stroke classification is crucial for clinical decision-making. Based on deep learning technology, this study proposed a lightweight and efficient GCDCNet model for intelligent classification of stroke CT images to improve the accuracy and efficiency of diagnosis. This study used a public stroke CT dataset, including 964 hemorrhagic strokes and 1551 ischemic strokes. Based on the GhostNet lightweight architecture, GCDCNet combines SCSE attention mechanism, dilated convolution, and channel shuffling for feature optimization to enhance the model's attention to the lesion area and improve classification performance. The model uses a cross entropy loss function and Adam optimizer, and is compared with ConvNeXt, EfficientNet, GhostNet, HRNet, and ViT. Grad-CAM visualization analysis is used to evaluate the model's discriminant basis. The experimental results show that GCDCNet has an accuracy of 99.26%, a recall rate of 99.33%, an F1 score of 99.13%, and a precision of 98.94% on the test set, with a parameter count of only 2.7 M, which is much lower than other mainstream deep learning models. The AUC-ROC curve analysis shows that GCDCNet has the best category discrimination ability (AUC = 0.9996), and the Grad-CAM visualization analysis shows that the model discrimination area is highly consistent with the clinician's focus area. In contrast, ConvNeXt and ViT performed poorly in this task. GCDCNet demonstrates superior accuracy, lightweight, and good generalization ability in the intelligent classification task of stroke CT images through lightweight design, feature optimization, and enhanced interpretability. This model can be used to assist clinicians in early stroke screening, emergency triage, and telemedicine applications, improve stroke diagnosis efficiency, and optimize medical resource allocation.