Unsupervised Learning for CT Image Segmentation via Adversarial Redrawing

Youyi Song, Teng Zhou, Jeremy Yuen-Chun Teoh, Jing Zhang, Jing Qin

Abstract

We propose a novel adversarial learning framework for unsupervised training of CNNs in CT image segmentation. It is motivated by difficulties in collecting voxel-wise annotations, which is laborious, time-consuming and expensive. It is conceptually simple, allowing us to train an effective segmentation network without any human annotation. Specifically, we design the generator with a CNN producing the segmentation results and a decoder redrawing the CT volume based on the segmentation results. The CNN is then implicitly trained in the adversarial learning framework where a discriminator gradually enforcing the generator to generate CT volumes whose distribution well matches the distribution of the training data. We further propose two constrains as regularization schemes for the training procedure to drive the model towards optimal segmentation by avoiding some unreasonable results. We conducted extensive experiments to evaluate the proposed method on a famous publicly available dataset, and the experimental results demonstrate the effectiveness of the proposed method.
BibTeX
@inproceedings{song2020unsupervised,
  author = {Youyi Song and Teng Zhou and Jeremy Yuen-Chun Teoh and Jing Zhang and Jing Qin},
  title = {{Unsupervised Learning for CT Image Segmentation via Adversarial Redrawing}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2020},
  doi = {10.1007/978-3-030-59719-1\_31},
}

Loading related content...

Figure