Cross Pseudo Method for Semi-supervised Abdominal Organ Segmentation

Junhao Dong, Tianyu Xiao, Ruining Zhao, Ziqi He, Xu Ji

Keywords: Multi-organ Segmentation, Cross Pseudo Supervision, Semisupervised Learning

TL;DR: A cross pseudo based method is proposed for semi-supervised abdominal organ segmentation.

Abstract: Abdominal multi-organ segmentation is of great significant for preoperative treatment planning. At present, there are many public abdominal datasets and deep learning based segmentatiomethods have been proposed. However, the problem of polycentric and spatio-temporal inefficiency still remain unsolved. Meanwhile, expensive costs of labeling and lack of labeled data are also serious problems of this field. In this work, with a small amount of labeled CT images and large number of unlabeled data, we propose a novel Cross Pseudo based semi supervision method, whose two branches can generate pseudo-labels to supervise each other. For quantitative evaluation on the FLARE2022 validation cases, this method achieves the DSC of 0.80, NSD of 0.75 within merely 20s for inference per image.It demonstrates the robustness and generalization of our method.