Semi-supervised Augmented 3D-CNN for FLARE22 Challenge

Zining Chen, Tianyi Wang

Keywords: Abdominal organ segmentation, Semi-supervised learning

Abstract: Abdominal organ segmentation has been used in many important clinical applications, however, cases with accurate labels require huge manual labour and financial resources. As a potential alternative, semi-supervised learning can explore useful information from unlabeled cases, with only few labeled cases involved. Therefore, we propose our baseline model using augmented 3D-UNet and adopt semi-supervised method--Mean Teacher, to make quantitative evaluation on the FLARE2022 validation cases. Our method achieves average dice similarity coefficient (DSC) of 62.16$\%$, Normalized Surface Distance (NSD) of 62.27$\%$, running time of 9.58s, and AUC of GPU and CPU is only 7424 and 199 respectively, which surpasses almost all other teams on resource consumption, demonstrating the effectiveness of our methods.