| Summary: | Annotating a large amount of medical imaging data thoroughly for training purposes can be expensive, particularly for medical image segmentation tasks; whereas obtaining scribbles, a less precise
form of annotation, is more feasible for clinicians. Nevertheless, training semantic segmentation networks with limited-signal supervision remains a technical challenge. In this paper, we present an innovative
scribble-supervised image segmentation via densely ensembling dense
pseudos called Collaborative Hybrid Networks(CHNets), which consists
of groups of CNN- and ViT-based segmentation networks. A simple yet
efficient densely collaboration scheme is introduced to ensemble dense
pseudo label to expand dataset allowing full-signal supervision. Additionally, internal consistency and external consistency training among
networks are proposed to ensure that each network is beneficial to the
other, resulting in a significant improvement. Our experiments on a public MRI benchmark dataset demonstrate that our proposed approach
outperforms other weakly-supervised methods on various metrics.
|
|---|