Unpaired Multi-contrast MR ImageSynthesis Using GenerativeAdversarial Networks

Abstract

Magnetic Resonance Imaging (MRI) has been established as an important diagnostic tool for research and clinical purposes. Multicontrast scans can enhance the accuracy for many deep learning algorithms. However, these scans may not be available in some situations. Thus, it is valuable to synthetically generate non-existent contrasts from the available one. Existing methods based on Generative Adversarial Networks (GANs) lack the freedom to map one image to multiple contrasts using only a single generator and discriminator, hence, requiring training of multiple models for multi-contrast MR synthesis. We present a novel method for multi-contrast MR image synthesis with unpaired data using GANs. Our method leverages the strength of Star-GAN to translate a given image to n contrasts using a single generator and discriminator. We also introduce a new generation loss function, which enforces the generator to produce high-quality images which are perceptually closer to the real ones and exhibit high structural similarity as well. We experiment on IXI dataset to learn all possible mappings among T1-weighted, T2-weighted, Proton Density (PD) weighted and Magnetic Resonance Angiography (MRA) images. Qualitative and quantitative comparison against baseline method shows the superiority of our approach.

Citation Style:    Display Alpha ACM SIGGRAPH IEEEtran Publication

Unpaired Multi-contrast MR ImageSynthesis Using GenerativeAdversarial Networks. Muhammad Sohail, Muhammad Naveed Riaz, Jing Wu, Chengnian Long, Shaoyuan Li. Simulation and Synthesis in Medical Imaging, 4th International Workshop, SASHIMI 2019, 11827, pp. 22–31, 2019.Muhammad Sohail, Muhammad Naveed Riaz, Jing Wu, Chengnian Long, and Shaoyuan Li. Unpaired multi-contrast mr imagesynthesis using generativeadversarial networks. Simulation and Synthesis in Medical Imaging, 4th International Workshop, SASHIMI 2019, 11827:22–31, October 2019.Sohail, M., Naveed Riaz, M., Wu, J., Long, C., and Li, S. 2019. Unpaired multi-contrast mr imagesynthesis using generativeadversarial networks. Simulation and Synthesis in Medical Imaging, 4th International Workshop, SASHIMI 2019 11827 (Oct.), 22–31.M. Sohail, M. Naveed Riaz, J. Wu, C. Long, and S. Li, “Unpaired multi-contrast mr imagesynthesis using generativeadversarial networks,” Simulation and Synthesis in Medical Imaging, 4th International Workshop, SASHIMI 2019, vol. 11827, pp. 22–31, Oct. 2019. Publisher Page BibTeX DOI