Abstract
Accurate and timely diagnosis of brain stroke is critical for effective clinical intervention and long-term patient outcomes. In this study, a novel deep learning-based framework for automated stroke diagnosis is proposed, utilizing an Auxiliary Branch Guided Swin Transformer with pseudo-segmentation supervision. The proposed architecture combines the hierarchical representation power of the Swin Transformer with a parallel auxiliary segmentation branch to enhance lesion-specific attention and spatial awareness. To address the scarcity of detailed annotations in clinical datasets, we employ pseudo-labels generated from bounding box–level supervision, enabling the model to learn lesion localization without full pixel-wise segmentation masks. The model was trained and validated on the ISLES 2024 dataset, which includes multimodal brain MRI scans. Quantitative results demonstrate that the proposed model achieves 94.6% accuracy, 94.3% precision, 94% recall, and an F1-score of 94%, outperforming existing CNN-based and transformer-based approaches. The auxiliary branch not only facilitates better feature refinement but also improves generalization by promoting regularization during training. This study highlights the effectiveness of transformer-based architectures in medical image analysis and introduces a practical solution for weakly-supervised stroke detection, offering a promising tool for clinical decision support and automated neuroimaging diagnostics.