Efficient Semi-Supervised Multi-Organ Segmentation Using Uncertainty Rectified Pyramid Consistency

Meng Han, Yijie Qu, Xiangde Luo

Keywords: Semi-supervised learning, Multi-organ segmentation, Uncertainty rectifying, Pyramid consistency

Abstract: To meet the problems that great dependence on fully annotated data and spatio-temporal inefficiency of remaining automatic multi-organ segmentation methods, an efficient semi-supervised framework with uncertainty rectified pyramid consistency regularization is introduced. Specifically, inspired by the fact that the predictions of the same input should be similar under different disturbance, we extend a backbone to produce predictions at different scale for unlabeled images and encourage them to be consistent. Since the multi-scale predictions have different resolution, directly encouraging them to be consistent may bring problems including lost of fine detail or model collapse. So a rectified scale-level uncertainty-aware module is introduced to enable the framework to gradually learn from reliable prediction regions. To deal with the domain gaps among multi-center datasets, a number of prepocessing methods are utilized, such as resampling the multi-center CT volumes to the same spacing and adjusting the window level and width. Quantitative evaluation on the FLARE2022 20 validation cases, this method achieves the average dice similarity coefficient (DSC) of 0.793 and average normalized surface distance (NSD) of 0.852.