A novel hybrid adaptive transformer framework with multihead self attention for stroke detection

Stroke detection using artificial intelligence (AI) continues to show promise in advancing medical imaging diagnostics, particularly in improving accuracy and accelerating clinical decision-making. This study presents VINCE-NETv1, a hybrid deep learning framework that integrates Vision Transformers (ViTs), Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) modules, and a meta-learning component to capture spatial, temporal, and global features from CT stroke images. Designed with architectural efficiency in mind, the model was trained and evaluated separately on three datasets: Near East University Hospital CT, the Core-Penumbra Acute Ischemic Stroke Dataset (CPAISD), and the Kaggle Brain Stroke CT dataset. To assess optimization performance, four optimizers-Adam, AdamW, Lookahead, and RMSProp-were used during training, each demonstrating high classification accuracy, with VINCE-NETv1 achieving up to 99.9% on CPAISD and ~ 100% on the Near East dataset. Generalization claims are limited strictly to datasets with confirmed patient-level separation-specifically the Near East and CPAISD datasets. The Kaggle dataset, lacking metadata for patient disambiguation, was treated solely as an exploratory benchmark. Preliminary interpretability was demonstrated using Grad-CAM visualizations, suggesting alignment with stroke-relevant regions. Approximate confidence intervals were computed from repeated runs to provide early insight into performance robustness, though formal statistical validation and k-fold cross-validation are planned in future work. While VINCE-NETv1 shows architectural potential for scalable deployment and clinical integration, current claims regarding generalization, real-time feasibility, and model interpretability are made cautiously with acknowledgment of existing limitations. Future research will address these aspects to evolve VINCE-NETv1 into a reliable and interpretable clinical decision-support tool.