Purpose: Conventional ways of synthetic-CT images (sCT) generation require high quality MR-CT pairs, which limits the possibility of exploiting the large amount of unpaired images. In this study, we developed an unsupervised machine learning model based on cycle-consistent generative adversarial networks (cycle-GAN) for the MR-based sCT generation and evaluated with the nasopharyngeal patient data.
Methods: CT and T1-weighted MR images were acquired from each of 157 nasopharyngeal patients at the treatment position on the same day. Eleven paired CT-MR images were randomly chosen as tested set and the remaining images were used as training set. An unsupervised learning model (cycle-GAN), containing two discriminators (patch-GAN) and two generators based on the deep residual learning was implemented and trained with unsupervised setting, i.e. not applying MR-CT registration for the training samples. The generated sCTs were compared to the ground true CT image using mean absolute error (MAE). The accuracy of the generated sCT images were compared that of the sCT images generated by a standard 2 dimensional 23-layers U-net trained with fully supervised setting, i.e. applying MR-CT registration on the training samples.
Results: The (MAE)s were (52.18 Â± 13.21) HU vs (63.33 Â± 18.11) HU (p = 0.011) within whole body, (173.99 Â± 45.6) HU vs (216.39 Â± 82.55) HU (p = 0.024) within bone (HU > 150) and (28.48 Â± 7.62) HU vs (33.03 Â± 6.84) HU (p = 0.031) within soft-tissue (150 > HU > -100) for the cycle-GAN vs the U-net, where the p was calculated using the student t-test.
Conclusion: Cycle-GAN trained with unsupervised setting outperformed the U-net trained with fully supervised setting significantly. Further study needs to improve the sCT image accuracy and investigate the dosimetric accuracy to use the generated sCT for the nasopharyngeal radiotherapy treatment planning.