<p>Compared with the fully supervised segmentation that highly relies on pixel-wise annotations, weakly supervised segmentation based on box labels can maximize the utilization of annotations from clinical practice. However, existing weakly supervised methods lack direct perception of the shape of segmentation targets and fail to balance shape and location constraints. In this paper, we propose a weakly supervised segmentation method based on shape-guided adaptive mutual training. Specifically, we design the pixel aggregation based superpixel generation module, which utilizes neural networks to generate more cohesive superpixel blocks from superpixel segmentation and selects appropriate ones to produce the so-called superpixel labels. Then, we propose the confidence-based cross-pseudo-supervision mechanism, which adaptively selects superpixel labels, pseudo-labels, and box labels based on their consistency with the actual lesion location, alleviating the illusionary network training under the shape-only guidance. Experimental results on two ultrasound datasets demonstrate that our method outperforms existing weakly supervised methods in segmentation performance and is comparable to fully supervised methods. The proposed method facilitates the integration of the deep learning methods into the practical medical workflows.</p>