Prototype Consistency Learning for Medical Image Segmentation by Cross Pseudo Supervision

Due to the acquisition of anatomical/pathological labels is expensive and time-consuming, semi-supervised semantic segmentation is commonly utilized in medical image analysis. Previous studies have overlooked the high similarity of the pixels in medical images, resulting in many models cannot effectively distinguish the pixels of different categories. A new semi-supervised semantic segmentation framework based on prototype learning is proposed in this paper. It contains a feature extractor and a superpixel-based graph convolutional network (GCN). Two consistency loss functions are proposed in our paper. The prototype cyclic consistency loss is utilized to incorporate explicit guidance of the labeled data; the cross pseudo supervised loss is applied to make full use of the context information of the unlabeled data. We evaluate the effectiveness of our proposed method on two classical public medical image datasets (MC and JSRT). On MC dataset, the predicted IoU of our method is 94.92 ±0.5% with only 25% annotated data; on JSRT dataset, the MIoU of our method reaches 89.51 ±0.37% (with 25% annotated data) and 90.98 ±0.4% (with 50% annotated data). Our proposed method outperforms most existing semi-supervised semantic segmentation methods, even exceeds the fully supervised semantic segmentation methods, and achieves high-precision semi-supervised semantic segmentation effectively.