Weakly Supervised Medical Image Segmentation Through Dense Combinations of Dense Pseudo-Labels

Annotating a large amount of medical imaging data thoroughly for training purposes can be expensive, particularly for medical image segmentation tasks. Instead, obtaining less precise scribble–like annotations is more feasible for clinicians. In this context, training semantic segmentation networks with limited-signal supervision remains a technical challenge. We present an innovative scribble-supervised approach to image segmentation via densely combining dense pseudo-labels which consists of groups of CNN– and ViT–based segmentation networks. A simple yet efficient dense collaboration scheme called Collaborative Hybrid Networks (CHNets) ensembles dense pseudo–labels to expand the dataset such that it mimics full-signal supervision. Additionally, internal consistency and external consistency training of the collaborating networks are proposed, so as to ensure that each network is beneficial to the others. This results in a significant overall improvement. Our experiments on a public MRI benchmark dataset demonstrate that our proposed approach outperforms other weakly-supervised methods on various metrics. The source code of CHNets, ten baseline methods, and dataset are available at https://github.com/ziyangwang007/CV-WSL-MIS.