Accurate medical image segmentation is very important for disease diagnosis and treatment planning. The existing segmentation networks have achieved good performance in some medical image segmentation tasks. However, due to the complex conditions of lesions in size, shape, texture and boundary irregularity, current methods still have limitations. In this paper, we propose a multi-scale deformable attention-based MRI image segmentation network DefA-Net, which can fully utilize the context information and explore the exact lesion boundaries. In order to enhance the learning ability of the network, the bounding boxof the Ground Truth (GT) focus profile is used as auxiliary information to guide feature extraction through different scales and angles. A high-level feature weighted sampling fusion decoder is put forward generating suitable decoding results according to the influence of different high-level features and the deformable attention together with the reverse attention mechanism isapplied to locate detailed boundary clues to improve the accuracy. We also apply DefA-Net to our lesion-detection datasetStrokeQD in transfer learning mode and good segmentation results are acquired. Experiments on ISLES and StrokeQD provethe high accuracy and applicability of the proposed algorithm.