Synthesizing whole-body CT images from Dixon MR images using tissue composition prior knowledge and a parametric physical model

Magnetic resonance imaging (MRI) has replaced computed tomography (CT) in various medical applications since the excellent tissue contrast of the MR images enables the clear visualization of soft tissues and lesions. However, clinically essential physical tissue parameters that were originally derived from CT images for medical physics applications cannot be obtained from MR images due to the lack of fundamental relationships between MR signals and the parameters. This study proposed a novel method for synthesizing CT images from Dixon MR images. The proposed method segmented human tissues into three classes with corresponding compositional sub-tissue pairs. Composition maps were calculated using the standard composition of the sub-tissues with the volume fractions of water and fat obtained from Dixon images. A parametric physical model was applied to synthesize CT images with the composition maps and spectrum characteristic parameters of a reference CT scanner. The performance of the proposed method was evaluated using whole-body CT and Dixon images and compared with two generative adversarial networks (GANs). Results indicated that the proposed method accurately synthesized CT images from Dixon images with performance superior to the GANs. Compared with the GANs, on average, the proposed method lowered the absolute Hounsfield unit difference by 22, elevated the peak signal-to-noise ratio by 4.6 dB, and increased the structural similarity by 4 %. In conclusion, the proposed CT image synthesis method can be used in various clinical medical applications involving MR image integration to improve the accuracy of various medical physics calculations.