A 2D U-NET combined model based on lesion size for automated stroke lesion segmentation