Semi-supervised medical image segmentation with diverse views via cross-pseudo supervision

Semi-supervised learning has attracted more and more attention in medical image segmentation as it alleviates reliance on high-cost annotated data. Meanwhile, existing semi-supervised methods that utilize few labeled data alongside a larger amount of unlabeled data are limited to scenarios where the labeled data comprises at least 10% of the total. However, the effectiveness of these methods significantly diminishes when the proportion of labeled data is extremely low, such as 1% or 5%. Concentrating on this vital yet infrequently addressed scenario, a new method called Diverse Views Cross-Pseudo Supervision (DVCPS) is proposed in this paper. Diverse views at both the data-level and feature-level for unlabeled data are constructed by this method. The Feature Dual Views Network (FDVNet) is introduced, comprising a shared encoder and two distinct decoders. FDVNet is designed to extract both consensus and complementary information from inconsistent views. The predictions generated by the student FDVNet and the pseudo-labels generated by the teacher FDVNet, each produced by different decoders for unlabeled data, are utilized for cross-supervision to acquire complementary information. DVCPS was validated on the ACDC dataset and the MSD Prostate dataset. Compared with other current state-of-the-art methods at specific proportions of labeled data, the best performance in semi-supervised segmentation was achieved by DVCPS. With over a doubling of Dice improvement on each dataset with only 1% labeled data, this demonstrates that DVCPS can effectively address the extreme imbalance of labeled data.