摘要
目的探讨基于无监督扩散模型的锥形束CT(C BCT)到CT图像合成技术在放疗中的应用价值,以降低患者辐射暴露并提升自适应放疗精度。方法构建小波高频引导扩散模型(WHGD)。基于Daubechies-4小波频域框架实现C BCT图像四级解耦,精确分离其低频结构和高频细节,确保合成CT的解剖完整性;创新设计Sobel算子-变分自编码器-生成对抗网络(Sobel-VAE-GAN)引导机制,通过梯度特征映射和共享潜在空间,在反向扩散中恢复CBCT的高频解剖细节;实现无监督条件下的跨域泛化,增强模型在临床放疗场景中的鲁棒性。模型选用公开和内部数据集(头部/盆腔/胸部合计231对)进行训练和验证,并按照7:1:2的比例划分为训练集、验证集和测试集。采用峰值信噪比(PSNR)、结构相似性指数(SSIM)和均方误差(MSE)评估图像质量,并与6种先进的基线模型进行对比。结果头部数据集,WHGD模型的PSNR为29.975,SSIM为0.939,MSE与最佳基线相当。相较于CycleGAN,PSNR提升6.1%,SSIM提升2.1%;相较于MUNIT,PSNR提升8.1%,SSIM提升2.4%;相较于扩散模型TPDM,PSNR提升42.0%,SSIM提升12.9%。盆腔数据集,WHGD的PSNR为28.171,SSIM为0.947,MSE为0.00136,相较于次优FGDM方法的PSNR提升5.7%,MSE下降24.0%。胸部数据集,WHGD的PSNR为25.175,SSIM为0.932,MSE为0.00222,优于所有对比方法。视觉评估显示,WHGD在金属置入物伪影抑制和软组织结构保真度方面表现优异。消融实验进一步证实了各模块的有效性和高性能特性。结论该方法为无监督C BCT到CT合成提供了高效范式,有望显著减少患者辐射风险并推进精准放疗的临床应用。
Objective To investigate the clinical utility of unsupervised diffusion model for synthesizing CT images from CBCT in radiotherapy,with the aim of reducing patient radiation exposure and enhancing adaptive radiotherapy precision.Methods This paper introduces the Wavelet-domain High-frequency Guided Diffusion model(WHGD).Based on the Daubechies-4 wavelet frequency domain framework for four-level decoupling of CBCT images,precisely separating lowfrequency structures and high-frequency details to ensure anatomical integrity in synthesized CT images;a novel Sobel-VAE-GAN guidance mechanism that leverages gradient feature mapping and shared latent space to accurately recover CBCT high-frequency anatomical details during reverse diffusion;and unsupervised cross-domain generalization to enhance robustness in clinical radiotherapy scenarios.The model was trained and validated using public and in-house datasets(a total of 231 head/pelvis/chest pairs),which were split into training,validation,and test sets at a 7∶1∶2 ratio.Image quality was assessed using peak signal-to-noise ratio(PSNR),structural similarity index(SSIM),and mean squared error(MSE),with comparisons against six state-of-the-art baseline models.Results On the head dataset,WHGD achieved a PSNR of 29.975,SSIM of 0.939,and MSE comparable to the top baseline.Relative to CycleGAN,PSNR improved by 6.1%and SSIM by 2.1%;compared to MUNIT,PSNR increased by 8.1%and SSIM by 2.4%;versus the diffusion-based TPDM,PSNR rose by 42.0%and SSIM by 12.9%.For the pelvis dataset,WHGD yielded a PSNR of 28.171,SSIM of 0.947,and MSE of 0.00136,surpassing the second-best FGDM method with a 5.7%PSNR gain and 24.0%MSE reduction.On the chest dataset,WHGD attained a PSNR of 25.175,SSIM of 0.932,and MSE of 0.00222,outperforming all comparators.Visual evaluations highlighted WHGD′s excellence in suppressing metal implant artifacts and preserving soft tissue fidelity.Ablation studies validated the efficacy and superior performance of individual modules.Conclusion This approach establishes an efficient paradigm for unsupervised CBCT-to-CT synthesis,holding promise for substantially mitigating patient radiation risks and advancing precision radiotherapy in clinical practice.
作者
申珊珊
邹建军
李春泉
范耀华
吕亚慧
李艳
SHEN Shanshan;ZOU Jianjun;LI Chunquan(Department of Oncology,Second Hospital of Jiaxing,Jiaxing,Zhejiang Province 314000,P.R.China)
出处
《临床放射学杂志》
北大核心
2026年第4期702-710,共9页
Journal of Clinical Radiology
基金
浙江省医药卫生科技计划项目(编号:2024KY444)。
关键词
CBCT到CT合成
高频引导
小波扩散模型
无监督学习
放射治疗
CBCT-to-CT synthesis
High-frequency guidance
Wavelet diffusion model
Unsupervised learning
Radiotherapy
