Magnetic resonance imaging plays a crucial role in the diagnosis and management of ischemic stroke. Accurate segmentation of stroke lesions holds significant clinical value in assisting the formulation of individualized interventional treatment plans and objectively assessing patient prognosis. To address the challenges of blurred and irregularly shaped ischemic stroke lesions with random locations, this study proposes a prior knowledge-guided and multi-level edge feature fusion shifted window Transformer-based U-Net with encoder representations (PMSwin UNETR). First, based on the distribution characteristics of stroke lesions, a lesion distribution probability map is generated to guide the segmentation network in focusing on areas prone to lesions. Second, a multi-level edge feature extraction module is employed to enrich edge features. Finally, a soft-clDice loss function is introduced to directly learn lesion boundaries, enhancing edge segmentation accuracy. The effectiveness of PMSwin UNETR was validated using the 2022 Ischemic Stroke Lesion Segmentation Challenge (ISLES2022) dataset, with results showing Dice similarity coefficient (DSC), 95% Hausdorff distance (HD95), and recall rates of 82.43%, 3.768 8, and 82.45%, respectively, outperforming other mainstream segmentation algorithms. This study demonstrates that the proposed PMSwin UNETR model effectively improves ischemic stroke lesion segmentation, providing important references and application value for precise clinical diagnosis and intelligent medical imaging research in stroke.