Separated Contrastive Learning for Organ-at-Risk and Gross-Tumor-Volume Segmentation with Limited Annotation

Automatic delineation of organ-at-risk (OAR) and gross-tumor-volume (GTV) is great significance for radiotherapy planning. However, it a challenging task to learn powerful representations accurate under limited pixel (voxel)-wise annotations. Contrastive learning at pixel-level can alleviate the dependency on annotations by dense from unlabeled data. Recent studies in this direction design various contrastive losses feature maps, yield discriminative features each map. pixels same map inevitably share semantics be closer than they actually are, which may affect discrimination lead unfair comparison other maps. To address these issues, we propose separated region-level scheme, namely SepaReg, core separate image into regions encode region separately. Specifically, SepaReg comprises two components: structure-aware separation (SIS) module an intra- inter-organ distillation (IID) module. The SIS proposed operate set rebuild guidance structural information. representation will learned via typical cross regions. On hand, IID tackle quantity imbalance as tiny organs produce fewer regions, exploiting intra-organ representations. We conducted extensive experiments evaluate model public dataset private datasets. experimental results demonstrate effectiveness model, consistently achieving better performance state-of-the-art approaches. Code available https://github.com/jcwang123/Separate_CL.