Semi-supervised Segmentation with Self-training Based on Quality Estimation and Refinement

摘要

Building a large dataset with high-quality annotations for medical image segmentation is time-consuming and highly depends on expert knowledge. Therefore semi-supervised segmentation has been investigated by utilizing a small set of labeled data and a large set of unlabeled data with generated pseudo labels, but the quality of pseudo labels is crucial since bad labels may lead to even worse segmentation. In this paper, we propose a novel semi-supervised segmentation framework which can automatically estimate and refine the quality of pseudo labels, and select only those good samples to expand the training set for self-training. Specifically the quality is automatically estimated in the view of shape and semantic confidence using variational auto-encoder (VAE) and CNN based network. And, the selected labels are refined in an adversarial way by distinguishing whether a label is the ground truth mask or not at pixel level. Our method is evaluated on the established neuroblas-toma(NB) and BraTS18 dataset and outperforms other state-of-the-art semi-supervised medical image segmentation methods. We can achieve a fully supervised performance while requiring ×4x less annotation effort.