Abstract: In semi-supervised medical image segmentation, current methods typically generate pseudo labels at a fixed scale and refine them without sufficient attention to boundary regions. Such ...