Due to the lack of expertise for medical image annotation, the investigation\nof label-efficient methodology for medical image segmentation becomes a heated\ntopic. Recent progresses focus on the efficient utilization of weak annotations\ntogether with few strongly-annotated labels so as to achieve comparable\nsegmentation performance in many unprofessional scenarios. However, these\napproaches only concentrate on the supervision inconsistency between strongly-\nand weakly-annotated instances but ignore the instance inconsistency inside the\nweakly-annotated instances, which inevitably leads to performance degradation.\nTo address this problem, we propose a novel label-efficient hybrid-supervised\nframework, which considers each weakly-annotated instance individually and\nlearns its weight guided by the gradient direction of the strongly-annotated\ninstances, so that the high-quality prior in the strongly-annotated instances\nis better exploited and the weakly-annotated instances are depicted more\nprecisely. Specially, our designed dynamic instance indicator (DII) realizes\nthe above objectives, and is adapted to our dynamic co-regularization (DCR)\nframework further to alleviate the erroneous accumulation from distortions of\nweak annotations. Extensive experiments on two hybrid-supervised medical\nsegmentation datasets demonstrate that with only 10% strong labels, the\nproposed framework can leverage the weak labels efficiently and achieve\ncompetitive performance against the 100% strong-label supervised scenario.\n