Objective To explore the synthesis of high-quality CT (sCT) from cone-beam CT (CBCT) using PE-CycleGAN for adaptive radiotherapy (ART) for nasopharyngeal carcinoma. Methods A perception-enhanced CycleGAN model "PE-CycleGAN" was proposed, introducing dual-contrast discriminator loss, multi-perceptual generator loss, and improved U-Net structure. CBCT and CT data from 80 nasopharyngeal carcinoma patients were used as the training set, with 7 cases as the test set. By quantifying the mean absolute error (MAE), peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), as well as the dose gamma pass rate and the relative dose deviations of the target area and organs at risk (OAR) between sCT and reference CT, the image quality and dose calculation accuracy of sCT were evaluated. Results The MAE of sCT generated by PE-CycleGAN compared to the reference CT was (56.89±13.84) HU, approximately 30% lower than CBCT's (81.06±15.86) HU (P<0.001). PE-CycleGAN's PSNR and SSIM were 26.69±2.41dB and 0.92±0.02 respectively, significantly higher than CBCT's 21.54±2.37dB and 0.86±0.05 (P<0.001), indicating substantial improvements in image quality and structural similarity. In gamma analysis, under the 2 mm/2% criterion, PE-CycleGAN's sCT achieved a pass rate of (90.13±3.75)%, significantly higher than CBCT's (81.65±3.92)% (P<0.001) and CycleGAN's (87.69±3.50)% (P<0.05). Under the 3 mm/3% criterion, PE-CycleGAN's sCT pass rate of (90.13±3.75)% was also significantly superior to CBCT's (86.92±3.51)% (P<0.001) and CycleGAN's (94.58±2.23)% (P<0.01). The mean relative dose deviation of the target area and OAR between sCT and planned CT was within ±3% for all regions, except for the Lens Dmax (Gy), which had a deviation of 3.38% (P=0.09). The mean relative dose deviations for PTVnx HI, PTVnd HI, PTVnd CI, PTV1 HI, PRV_SC, PRV_BS, Parotid, Larynx, Oral, Mandible, and PRV_ON were all less than ±1% (P>0.05). Conclusion PE-CycleGAN demonstrates the ability to rapidly synthesize high-quality sCT from CBCT, offering a promising approach for CBCT-guided adaptive radiotherapy in nasopharyngeal carcinoma.
PE-CycleGAN模型在传统CycleGAN[11]的基础上进行了创新性改进,如(图1)所示。模型主要由两个生成器(GCT to CBCT、GCBCT to CT)和两个判别器(DCT、DCBCT)组成,GCT to CBCT和GCBCT to CT均采用了基于U-Net[15]的改进结构。如(图1)左下方所示,生成器包含编码器、解码器和跨层连接。编码器通过下采样逐步提取多尺度语义特征,解码器则通过上采样恢复图像细节。跨层连接使编码器的特征图直接传递到解码器对应层,有效融合了局部和全局信息。此外,引入ResNet[16]式的残差块(图中的"Res Block"),有效缓解深层网络中的梯度消失问题。另一个改进是生成器引入多感知损失。图1左下方"Generator Architecture"所示,生成器的多元化的损失函数设计能够从像素级、结构级和感知级多个层面优化生成图像的质量。判别器也创新采用了双对比度损失机制(图1),DCT不仅区分真实CT和sCT,通过引入CBCT为负样本,强制性使得sCT特征推离CBCT特征(Push Away箭头),还将会使sCT的特征拉近真实CT的特征空间,增强了sCT与真实CT的相似性(Pull Close箭头)。这种双重约束显著增强了模型对不同图像域特征的辨识能力和生成能力,使得生成的sCT在保持CT图像特征的同时,有效减少了CBCT的噪声和伪影。判别器基于PatchGAN[17]结构,在局部图像块上进行判别和引导,进一步提高了生成图像的细节真实度和准确性。
RusanovB, HassanGM, ReynoldsM, et al. Deep learning methods for enhancing cone-beam CT image quality toward adaptive radiation therapy: a systematic review[J]. Med Phys, 2022, 49(9): 6019-54.
[11]
KidaS, Kaji, NawaK, et al. Visual enhancement of Cone-beam CT by use of CycleGAN[J]. Med Phys, 2020, 47(3): 998-1010.
[12]
LiuJW, YanH, ChengHL, et al. CBCT-based synthetic CT generation using generative adversarial networks with disentangled representation[J]. Quant Imaging Med Surg, 2021, 11(12): 4820-34.
[13]
WangTH, LeiY, FuYB, et al. Machine learning in quantitative PET: a review of attenuation correction and low-count image reconstruction methods[J]. Phys Med, 2020, 76: 294-306.
RonnebergerO. Invited talk: U-net convolutional networks for biomedical image segmentation[M]//Informatik aktuell. Berlin, Heidelberg: Springer Berlin Heidelberg, 2017: 3.
[16]
HeKM, ZhangXY, RenSQ, et al. Deep residual learning for image recognition[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). June 27-30, 2016. Las Vegas, NV, USA. IEEE, 2016: 770-8.
[17]
IsolaP, ZhuJY, ZhouT, et al. Image-to-image translation with conditional adversarial networks[J]. IEEE TPAMI, 2021, 43(12): 4254-67.
[18]
WangR, WuZX, WengZJ, et al. Cross-domain contrastive learning for unsupervised domain adaptation[J]. IEEE Trans Multimedia, 2023, 25: 1665-73.
[19]
WangZ, BovikAC, SheikhHR, et al. Image quality assessment: from error visibility to structural similarity[J]. IEEE Trans Image Process, 2004, 13(4): 600-12.
[20]
GalićI, HabijanM, LeventićH, et al. Machine learning empowering personalized medicine: a comprehensive review of medical image analysis methods[J]. Electronics, 2023, 12(21): 4411.
[21]
HuynhE, HosnyA, GuthierC, et al. Artificial intelligence in radiation oncology[J]. Nat Rev Clin Oncol, 2020, 17(12): 771-81.
[22]
WendlingM, MorrowA, HoggarthM, et al. An efficient protocol for radiotherapy quality control with machine learning[J]. Med Phys, 2020, 47(4): 1526-34.
[23]
LeiY, TangXY, HigginsK, et al. Learning-based CBCT correction using alternating random forest based on auto-context model[J]. Med Phys, 2019, 46(2): 601-18.
[24]
JiangJ, SharfoAWM, MakRH, et al. Development and validation of an MRI‐only synthetic CT generation method using cycle‐consistent generative adversarial networks for prostate radiotherapy[J]. Med Phys, 2021, 48(1): 416-29.
[25]
ThummererA, ZaffinoP, MeijersA, et al. Comparison of CBCT based synthetic CT methods suitable for proton dose calculations in adaptive proton therapy[J]. Phys Med Biol, 2020, 65(9): 095002.
[26]
PulliamKB, HuangJY, HowellRM, et al. Comparison of 2D and 3D gamma analyses[J]. Med Phys, 2014, 41(2): 021710.
LeiY, HarmsJ, WangTH, et al. MRI-only based synthetic CT generation using dense cycle consistent generative adversarial networks[J]. Med Phys, 2019, 46(8): 3565-81.
[29]
LiuY, LeiY, WangY, et al. Evaluation of a deep learning-based synthetic CT generation method for MRI-only breast radiotherapy[J]. Phys Med Biol, 2020, 65(8): 085020.
[30]
MasperoM, SavenijeMHF, DinklaAM, et al. Dose evaluation of fast synthetic-CT generation using a generative adversarial network for general pelvis MR-only radiotherapy[J]. Phys Med Biol, 2018, 63(18): 185001.
[31]
KimJ, ParkS, YuH, et al. Deep learning-based synthetic CT generation from MR images for PET attenuation correction: A systematic review and meta-analysis[J]. IEEE T Radiat Plasma, 2022, 6(3): 273-287.