A Research On Synthesizing CT Images From Multisequence MR Images Of Patients With Nasopharyngeal Carcinoma Based On Generative Adversarial Networks

Nasopharyngeal carcinoma(NPC)is a common malignant tumor in the human nasopharynx.It is of great clinical significance to synthesize CT(computed tomography)images directly from MR(Magnetic resonance)images of patients with NPC and make radiotherapy plan,which can not only avoid the radiation injury caused by CT scanning,but also simplify the workflow and reduce the error and uncertainty caused by cross modal registration,which has important clinical significance.MR images with three sequences of T1 W,T2W,T1 C and CT images of 40 patients with NPC were acquired from the Sun Yat-sen University Cancer Center.The primary tumors were delineated by experienced doctors on MR images.All subjects were randomly divided into training sets and independent test sets with the ratio of 3:1.According to the characteristics of multi-sequence MR images,a multi-sequence feature deep fusion(MSFDF)strategy based on ASPP(Atrous spatial pyramid pooling)structure and attention mechanism was proposed to construct a multi-sequence feature deep fusion generator.Two generative adversarial networks including image-paired Pix2pix-based network and imageunpaired Cycle GAN-based network adopted multi-sequence fusion scheme and then were constructed to generate synthetic CT(s CT)images from multi-sequence MR images.The experimental results showed that the MAE(Mean absolute error),MSE(Mean square error)and PSNR(Peak signal to noise ratio)of s CT images synthesized by Pix2 pixbased multi-sequence fusion generative model were 94.76 HU,41650.98 HU and 26.35 d B,respectively.The MAE,MSE and PSNR of s CT synthesized by Cycle GAN-based multisequence fusion generative model were 122.40 HU,60782.58 HU and 24.49 d B,respectively..Furthermore,the quality of primary tumor region in s CT images was evaluated.The MAE,MSE and PSNR of s CT images synthe21 sized by Pix2pix-based multi-sequence fusion generative model were 69.89 HU,20670.67 HU and 29.54 d B,respectively.The Dice coefficient was 0.842,the Hausdorff distance was 4.86 mm,whereas the MAE,MSE and PSNR of s CT images synthesized by Cycle GAN-based multi-sequence fusion generatative model were 92.64 HU,29363.89 HU and 28.61 d B respectively,Dice coefficient was 0.787,Hausdorff distance was 5.34 mm.This study used multi-sequence MR images of NPC patients to construct two multisequence fusion generative models applying Pix2 pix and Cycle GAN architecture.Experimental results demonstrated that the proposed generative models could generate high quality s CT images from multi-sequence MR images of NPC patients.Especially the pixel intensity and structure consistency of the primary tumor region of s CT images and real CT images were verified.This study provided an effective strategy and method for generating high-quality s CT images from multi-sequence MR images,which exhibited significant potentials in clinical application.