Accurate medical image segmentation plays a pivotal role in diagnostic imaging and computer-aided intervention, particularly in delineating ischemic stroke lesions. Efficient and precise identification of these regions not only supports timely clinical decision-making but also helps secure the critical therapeutic window, thereby reducing the risk of irreversible neurological impairment. Nevertheless, many existing approaches rely on single-scale feature extraction, which limits their ability to capture multi-scale contextual information. In addition, the scarcity of annotated medical data highlights the importance of developing lightweight segmentation models. To address these challenges, we present a compact vision transformer–based framework capable of learning semantic representations across multiple scales. Evaluations on the ATLAS and ISLES2022 datasets indicate that the proposed method achieves high accuracy and robust lesion segmentation performance, underscoring its clinical value in stroke treatment.