Abstract
Accurate segmentation of stroke lesions in brain magnetic resonance imaging (MRI) is critical for early diagnosis and effective intervention. Existing convolutional neural networks (CNNs) have shown promising results but often struggle with global contextual reasoning and generalization in the presence of small, diffuse, or anatomically variable lesions. To address these limitations, we introduce a novel segmentation framework that integrates a Swin Transformer backbone with an auxiliary supervision mechanism based on bounding box-derived pseudo masks. Unlike prior transformer-based models that rely solely on end-to-end attention, our method introduces intermediate supervision via an auxiliary branch, which guides early layers to focus on lesion-relevant regions using weak annotations. This dual-path strategy enhances spatial representation learning while mitigating the annotation burden typically required for full supervision. Evaluated on the ISLES 2024 dataset, one of the most challenging benchmarks for ischemic lesion segmentation, the proposed model achieves superior performance in dice similarity, precision, and recall when compared to recent state-of-the-art CNN and vision transformer architectures. Qualitative results further highlight its robustness in capturing diverse lesion morphologies. By combining weak supervision with transformerbased learning, our approach contributes a scalable and annotation-efficient solution to neuroimaging, advancing the field of automated stroke diagnosis with improved accuracy and clinical feasibility.