Cascade Dual-decoders Network for Abdominal Organs Segmentation

Ershuai Wang, Yaliang Zhao, Yajun Wu

Keywords: Semi-Supervised Learning, Segmentation, nnUNet

Abstract: In order to make full use of unlabeled images, we developed a pseudo-label based localization-to-segmentation framework for efficient abdominal organs segmentation. To reduce the target region, we locate the abdominal by a U-Net, then we train a fine organ segmentation model, which reduce the maximum usage of RAM memory. Segmentation with Dual-decoders is designed to improve the stability and cross supervise each other by pseudo labels. We also propose a class-weighted loss to pay more attention on the small organs like gallbladder, pancreas, which improve the mean performance. Finally, we test the models on the public validation, the total running time for the 50 CT images is 6676 seconds, the mean DSC is 0.8830 and the mean NSD is 0.9189.