Brain MRI synthesis based on dual-generator generative adversarial network
Xianbiao Bai, Shunbo Hu, Jitao Li, Lei Wang, Dezhuang Kong · 2022
Various contrast Magnetic Resonance Imaging (MRI) of can increase the available information in clinical diagnosis. However, due to the time limit, cost, and the patient cooperation in the scanning process, some contrast images may not be obtained in time. Moreover, some images may be damaged during the sampling process. Therefore, the synthesis model is designed to obtain the missed contrast image from the existing image, which is beneficial to improve the effect of diagnosis. Hence, we propose a new brain MRI synthesis method based on generative adversarial networks (GAN). This method leverages a coarse-to-fine dual network structure based on the generative adversarial network and uses FID to make T2 weighted images generate more reliable T1 weighted images. We conducted experiments on 2730 T1 and T2 weighted images, and used structural similarity, peak signal-to-noise ratio, and mean square error as evaluation indicators. Quantitative and qualitative experimental results show that our method is superior to the original one.