Synthetic CT generation from NPC MRI using Transformer-based generative adversarial network

Fanghua LI; Shouliang DING; Yongbao LI; Biaoshui LIU; Li CHEN; Xiaoyan HUANG; Hongdong LIU

Return

Synthetic CT generation from NPC MRI using Transformer-based generative adversarial network

VernacularTitle:基于Transformer的生成对抗网络用于鼻咽癌MRI生成伪CT
Author: Fanghua LI ¹ ; Shouliang DING ¹ ; Yongbao LI ¹ ; Biaoshui LIU ¹ ; Li CHEN ¹ ; Xiaoyan HUANG ¹ ; Hongdong LIU ¹
Author Information

1. 中山大学肿瘤防治中心放疗科/华南肿瘤学国家重点实验室/肿瘤医学协同创新中心,广东广州 510060
Publication Type:Journal Article
Keywords: nasopharyngeal carcinoma; magnetic resonance imaging; synthetic computed tomography; Transformer; generative adversarial network
From: Chinese Journal of Medical Physics 2025;42(6):701-707
CountryChina
Language:Chinese
Abstract: Objective To compare the performance of two different deep learning models,VTcGAN and Pix2pix,in generating synthetic computed tomography(sCT)from magnetic resonance imaging(MRI)of nasopharyngeal carcinoma(NPC),and to evaluate their accuracies in treatment planning dose calculations.Methods MRI and CT images as well as treatment planning data of 115 NPC patients were retrospectively selected,and paired dataset was obtained through rigid image registration,with 105 cases as the training set,and the remaining 10 cases as the test set.Two kinds of models,namely Pix2pix model based on conventional convolutional neural network and the improved VTcGAN model based on Transformer network,were constructed with the consistent structure except for the bottleneck network in the generator.The generated sCT images(sCTPix2pix and sCTVTcGAN)were assessed in terms of image quality,intensity value and dosimetric differences.Results For the cases in test set,the mean error,mean absolute error,peak signal-to-noise ratio,and structural similarity index between the ground truth CT images and the sCTPix2pix were(-0.86±12.42)HU,(40.77±3.06)HU,(33.45±0.62)dB,and 0.928±0.013,respectively;and those between the ground truth CT images and the sCTVTcGAN were(-1.10±8.56)HU,(37.40±2.08)HU,(34.33±0.45)dB,and 0.936±0.009,respectively.For the criterion of 1 mm/1%,the averaged gamma passing rates of sCTPix2pix and sCTVTcGAN were(96.62±1.08)%and(96.88±0.99)%at a dose threshold of 10%,(94.31±1.03)%and(94.72±0.91)%at a dose threshold of 50%,(84.62±1.74)%and(86.06±1.41)%at a dose threshold of 80%,respectively.Conclusion The proposed VTcGAN model is superior to the traditional Pix2pix model in terms of accuracy in generating sCT from MRI of NPC,and it can fulfill the requirements for dose calculation in the MRI-Only workflow.