Semi-Supervised Cardiac Image Segmentation Using a Bicycle VAE with Cross Prior Attention
Shaojie Li, Yifan Zhang, Xuan Yang · 2022 IEEE International Conference on Bioinformatics and Biomedicine (BIBM) · 2022
This paper proposes a semi-supervised learning framework for cardiac image segmentation based on a bicycle Variational Auto-Encoder (biVAE) architecture by embedding a Prior Transformer and a conditional generative network. Specifically, to enhance the presentation of extracted features, a Transformer with cross attention to prior is embedded in the encoder of our framework, where the global context and anatomical prior are modeled at each layer of the Prior Transformer’s decoder to fuse multi-scale features. A biVAE is proposed by employing a conditional generative network, which maps the latent space to the image and maps the image to the latent space again. It makes the encoder of biVAE serve as a good feature representation for cardiac image segmentation. Many experiments on public cardiac MR datasets show that by introducing prior to the Transformer, more powerful features can be extracted in terms of representation, improving our semi-supervised framework’s predictive performance. Furthermore, the proposed semi-supervised framework achieves competitive results with state-of-the-art supervised learning-based networks.