Multi-Sequence Fusion Network via Single- Sequence CycleGANs for Improved Synthetic CT in Nasopharyngeal Carcinoma Treatment Planning
Yimei Liu,
Meining Chen,
Jun Zhang,
Yixuan Wang,
Huikuan Gu,
Chong Zhao,
Zhenyu Qi,
Xiaowu Deng,
Shuyu Wu,
Yinglin Peng
Affiliations
Yimei Liu
Department of Radiation Oncology, State Key Laboratory of Oncology in South China, Guangdong Key Laboratory of Nasopharyngeal Carcinoma Diagnosis and Therapy, Guangdong Provincial Clinical Research Center for Cancer, Sun Yat-sen University Cancer Center, Guangzhou, China
Meining Chen
Department of Radiation Oncology, State Key Laboratory of Oncology in South China, Guangdong Key Laboratory of Nasopharyngeal Carcinoma Diagnosis and Therapy, Guangdong Provincial Clinical Research Center for Cancer, Sun Yat-sen University Cancer Center, Guangzhou, China
Jun Zhang
Department of Radiation Oncology, State Key Laboratory of Oncology in South China, Guangdong Key Laboratory of Nasopharyngeal Carcinoma Diagnosis and Therapy, Guangdong Provincial Clinical Research Center for Cancer, Sun Yat-sen University Cancer Center, Guangzhou, China
Department of Radiation Oncology, State Key Laboratory of Oncology in South China, Guangdong Key Laboratory of Nasopharyngeal Carcinoma Diagnosis and Therapy, Guangdong Provincial Clinical Research Center for Cancer, Sun Yat-sen University Cancer Center, Guangzhou, China
Huikuan Gu
Department of Radiation Oncology, State Key Laboratory of Oncology in South China, Guangdong Key Laboratory of Nasopharyngeal Carcinoma Diagnosis and Therapy, Guangdong Provincial Clinical Research Center for Cancer, Sun Yat-sen University Cancer Center, Guangzhou, China
Chong Zhao
Department of Radiation Oncology, State Key Laboratory of Oncology in South China, Guangdong Key Laboratory of Nasopharyngeal Carcinoma Diagnosis and Therapy, Guangdong Provincial Clinical Research Center for Cancer, Sun Yat-sen University Cancer Center, Guangzhou, China
Zhenyu Qi
Department of Radiation Oncology, State Key Laboratory of Oncology in South China, Guangdong Key Laboratory of Nasopharyngeal Carcinoma Diagnosis and Therapy, Guangdong Provincial Clinical Research Center for Cancer, Sun Yat-sen University Cancer Center, Guangzhou, China
Department of Radiation Oncology, State Key Laboratory of Oncology in South China, Guangdong Key Laboratory of Nasopharyngeal Carcinoma Diagnosis and Therapy, Guangdong Provincial Clinical Research Center for Cancer, Sun Yat-sen University Cancer Center, Guangzhou, China
Shuyu Wu
Department of Radiation Oncology, Affiliated Cancer Hospital and Institute of Guangzhou Medical University, Guangzhou, China
Department of Radiation Oncology, State Key Laboratory of Oncology in South China, Guangdong Key Laboratory of Nasopharyngeal Carcinoma Diagnosis and Therapy, Guangdong Provincial Clinical Research Center for Cancer, Sun Yat-sen University Cancer Center, Guangzhou, China
To investigate the effect of different MR sequences on the accuracy of Cycle-consistent Generative Adversarial Network (CycleGAN) based synthetic CT (sCT) generation in nasopharyngeal carcinoma (NPC). In this work, three sequences of MR, included T1-weighted imaging (T1WI), T2-weighted imaging (T2WI), and T1 contrast-enhanced weighted imaging (T1WIC), and planning CT (PCT) images of 151 patients with NPC were collected. The number of training, verification, and test sets were 108, 16, and 27, respectively. Three unsupervised CycleGAN-based models were trained using different sequences (single channels) as inputs, and the synergistic fusion model were used multiple channels. To assess the precision of these models, we evaluated key metrics such as mean error (ME), mean absolute error (MAE), structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), and dose distribution, comparing the PCT with the sCT generated by each model. Overall, The SCTT2 image generated by T2WI model achieved superior results than those of T1WIC and T1WI for a single-sequence model. Compared to the SCTT2, the quality of the SCTFusion image generated by the fusion model was further improved, with significant differences in the MAE, SSIM, and PNSR metrics (P<0.05), the average MAE, SSIM, and PNSR in the body were ( $112.97\pm 9.69$ vs. $90.99\pm 9.64$ , P<0.05) HU, ( $0.86\pm 0.02$ vs. $0.89\pm 0.02$ , P<0.05), and ( $21.66\pm 0.64$ vs. $23.28\pm 0.87$ , P<0.05), respectively. Meanwhile, the average gamma passing rates (3%, 3 mm) and the average absolute dose discrepancies were 99.45% $\pm ~1.05$ % and 0.73%±0.73% for the fusion model. In conclusion, our findings reveal that the CycleGAN model, particularly when employing fusion MR sequences as input, offers the highest accuracy in synthetic CT generation. Notably, T2WI images stand out as a viable option for sCT prediction in clinical settings where acquisition sequences or times are limited.
