Semi-Supervised 3D U-Net learning based on Meta Pseudo Labels

Xiao Chu D, Zhuo Chen, Dan Li, Haoyu Li, Rashid khan, Jinyu Tian, Weiguo Xie, Liyilei Su

Keywords: Meta Pseudo Labels, Semi-supervised learning, One-step gradient, 3D U-Net

TL;DR: using Meta Pseudo Labels to segment medical images

Abstract:

Deep learning models have demonstrated promising performance for segmenting medical images and are significantly dependent
on a huge amount of well-annotated data. However, it is difficult to get a large amount of data, particularly in clinical practices. Likewise, high-performance deep learning models have an enormous model size, restricting their use in actual applications. In order to reduce the burden of both expensive annotations and computational expenses, we designed the semi-supervised knowledge-based method on top of 3D U-Net and Meta Pseudo Labels. We train the teacher network with labeled data to generate the pseudo labels. And then we train the student network on the pseudo labels, and give the training feedback to the teacher network. The student network on FLARE2022 grand challenge Dataset achieved 81.19 % of DSC and 85.20% of NSD. As for the network inference speed, it needs 50.59s for a single case.