Unlabeled Abdominal Multi-organ Image Segmentation Based on Semi-supervised Adversarial Training Strategy

Abstract

AbstractThe unlabeled images are helpful to generalize segmentation models. To make full use of the unlabeled images, we develop a generator-discriminator training pipeline based on the EfficientSegNet, which has achieved the best performance and efficiency in previous FLARE 2021 challenge. For the generator, a coarse-to-fine strategy is used to produce segmentations of abdominal organs. Then the labeled image and the ground truth are applied to optimize the generator. The discriminator receives the original unlabeled image or the relevant noised image, together with their generated segmentation results to determine which segmentation is better for the unlabeled image. After the adversarial training, the generator is used to segment the unlabeled images. On the FLARE 2022 final testing set of 200 cases, our method achieved an average dice similarity coefficient (DSC) of 0.8497 and a normalized surface dice (NSD) of 0.8915. In the inference stage, the average inference time is 11.67 s per case, and the average GPU (MB) and CPU (%) consumption per case are 311 and 225.6, respectively. The source code is freely available at https://github.com/Yuanke-Pan/Adversarial-EfficientSegNet.KeywordsUnlabeled image segmentationSemi-supervised learningAdverarial training