Semi-Supervised Segmentation-Guided Tumor-Aware Generative Adversarial Network for Multi-Modality Brain Tumor Translation

Chuan Huang, Jia Wei, Rui Li

Keywords: brain tumor translation, multi-modality

TL;DR: Tumor-aware multi-modality brain tumor translation.

Abstract: Multi-modality brain tumor images are widely used for clinical diagnosis since they can provide complementary information. Yet, due to considerations such as time, cost, and artifacts, it is difficult to get fully paired multi-modality images. Therefore, most of the brain tumor images are modality-missing in practice and only a few are labeled, due to a large amount of expert knowledge required. To tackle this problem, multi-modality brain tumor image translation has been extensively studied. However, existing works often lead to tumor deformation or distortion because they only focus on the whole image. In this paper, we propose a semi-supervised segmentation-guided tumor-aware generative adversarial network called $S^3TAGAN$, which utilizes unpaired brain tumor images with few paired and labeled ones to learn an end-to-end mapping from source modality to target modality. Specifically, we train a semi-supervised segmentation network to get pseudo labels, which aims to help the model focus on the local brain tumor areas. The model can synthesize more realistic images using pseudo tumor labels as additional information to help the global translation. Experiments show that our model achieves competitive results on both quantitative and qualitative evaluations. We also verify the effectiveness of the generated images via the downstream segmentation tasks.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Deep Learning and representational learning