Synthetic CT Generation From Multi-Sequence MR Images for Head and Neck MRI-Only Radiotherapy via Cycle-Consistent Generative Adversarial Network

Abstract

Synthetic CT (sCT) generation with MR image plays an essential role in MR-only radiotherapy, which can be used for electron density estimation and dose calculation in treatment planning. Currently, the reliability of deep learning-based methods for sCT generation depends on two aspects: the MR-CT registration errors, which can eliminate by a cycle-consistent generative adversarial network (CycleGAN) with unpaired training data, and the different anatomy information and contract divergence of various tissue in different sequences MRI which will significantly affect the nonlinear mapping of MR-CT. The purpose of this study is to investigate the effect of different sequence MRI by evaluating the image quality of corresponded sCTs.Multi-sequence MR images (T1, T2, T1C) and plan CT images of 151 patients with nasopharyngeal carcinoma were acquired on the same day in the radiation treatment position. 126 cases were randomly selected as the training set with the same preprocessing. Three unsupervised CycleGAN-based models were respectively trained with different sequence MRI with Architecture, which consists of 2 ResUnet-based generators and 2 Patch-GAN-based discriminators. The remaining 25 CT-MR cases with non-rigid registration were selected as the independent paired testing set. The mean error (ME) and mean absolute error (MAE) between plan CT and sCT of each model were used to measure the CT value estimation accuracy, which was calculated within air (Hounsfield Unit, HU < 100), soft tissue (-100 < HU < 150), bone (HU > 150) and whole body, respectively. Structural similarity index (SSIM) was applied to evaluate the similarity between error (plan CT and sCT.The MEs calculated between plan CT and T1, T2, T1C-based sCT images, were (-0.62 ± 3.06) vs. (-1.81 ± 3.13) vs. (0.13 ± 2.56) within air, (154.92 ± 87.93) vs. (173.27 ± 97.62) vs. (262.78 ± 91.74) within soft tissue, (1363.99 ± 294.94) vs. (337.21 ± 50.4) vs. (661.63 ± 168.43) within bone, (22.12 ± 29.47) vs. (17.05 ± 10.51) vs. (35.28 ± 20.75) within the whole body. The MAEs correspondingly were (4.49 ± 3.13) vs. (4.6 ± 3.47) vs. (4.6 ± 3.47) within air, (186.41 ± 61.75) vs. (197.33 ± 58.62) vs. (293.56 ± 72.72) within soft tissue, (186.41 ± 61.75) vs. (197.33 ± 58.62) vs. (293.56 ± 72.72) within bone, (31.28 ± 28.1) vs. (27.64 ± 10.41) vs. (42.6 ± 23.82) within the whole body. The quantitative analysis between T1, T2, T1C-sCTs indicates that the T1C based generated model achieves higher accuracy than the other models in nasopharyngeal carcinoma.The T1C MRI can provide sufficient anatomy information and tissue contract in synthetic CT prediction of head and neck region, which offers a substantial potential for density estimation and dose calculation in radiotherapy. The generation method and evaluation will investigate synthetic CT prediction with more complex tissue in various body parts.