Neural Computing and Applications Программное обеспечение

ResNet-based deep learning architecture with attention mechanisms and Particle Swarm Optimization for brain stroke prediction using CT images

Ramalakshmi, K., Shukla, Vinod Kumar, Kumari, L. Krishna, Santhoshkumar, S. P., Pandiselvi, T., Krishnan, R. Santhana
Springer Nature · 2026-11-21 · DOI 10.1007/s00521-026-12632-9

Аннотация

Stroke is a life-threatening neurological disorder in which rapid and accurate diagnosis is essential for timely treatment and improved patient outcomes. Automated analysis of computed tomography (CT) images using deep learning has shown considerable potential for assisting stroke detection; however, achieving high diagnostic accuracy while effectively identifying clinically relevant image features remains challenging. This study proposes a ResNet-based deep learning architecture integrated with a Squeeze-and-Excitation (SE) attention mechanism and Particle Swarm Optimization (PSO) for automated brain stroke classification from CT images. The ResNet backbone extracts hierarchical image features, the SE attention mechanism emphasizes informative feature channels, and PSO optimizes model hyperparameters to improve classification performance. The proposed framework was evaluated on two publicly available brain stroke CT datasets. On the Kaggle Binary Brain Stroke CT Dataset, the model achieved 98.72% accuracy, with a ROC-AUC of 0.93 and a PR-AUC of 0.94. On the TEKNOFEST-2021 Brain Stroke CT Dataset, it achieved 98.19% accuracy, with a macro-average ROC-AUC of 0.986 and a macro-average PR-AUC of 0.973, across No Stroke, Ischemia and Bleeding classes. Comparative experiments further showed improved performance with the integration of SE attention and PSO over the baseline ResNet architecture. These results demonstrate promising classification performance on the evaluated public datasets; however, the absence of independent multi-center clinical validation and prospective evaluation limits conclusions regarding broader clinical generalizability and real-world clinical application.

Читать у издателя →