Cycle-consistent generative adversarial network (CycleGAN) has been widely used for cross-domain medical image systhesis tasks particularly due to its ability to deal with unpaired data. However, most CycleGAN-based synthesis methods can not achieve good alignment between the synthesized images and data from the source domain, even with additional image alignment losses. This is because the CycleGAN generator network can encode the relative deformations and noises associated to different domains. This can be detrimental for the downstream applications that rely on the synthesized images, such as generating pseudo-CT for PET-MR attenuation correction. In this paper, we present a deformation invariant model based on the deformation-invariant CycleGAN (DicycleGAN) architecture and the spatial transformation network (STN) using thin-plate-spline (TPS). The proposed method can be trained with unpaired and unaligned data, and generate synthesised images aligned with the source data. Robustness to the presence of relative deformations between data from the source and target domain has been evaluated through experiments on multi-sequence brain MR data and multi-modality abdominal CT and MR data. Experiment results demonstrated that our method can achieve better alignment between the source and target data while maintaining superior image quality of signal compared to several state-of-the-art CycleGAN-based methods.
CITATION STYLE
Wang, C., Papanastasiou, G., Tsaftaris, S., Yang, G., Gray, C., Newby, D., … MacGillivray, T. (2019). TPSDicyc: Improved Deformation Invariant Cross-domain Medical Image Synthesis. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 11905 LNCS, pp. 245–254). Springer. https://doi.org/10.1007/978-3-030-33843-5_23
Mendeley helps you to discover research relevant for your work.