Label-Efficient Hybrid-Supervised Learning for Medical Image Segmentation

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