基于稀疏重建锥形束CT和深度学习技术合成双能物质分解图

Synthesis of dual-energy material decomposition images based on sparse-view cone beam CT reconstruction and deep learning

目的利用治疗当日低剂量单能锥形束CT(CBCT),合成与治疗当日解剖结构一致的双能物质分解图像(MDI),为在线自适应放疗(ART)和剂量重建等临床应用场景提供定量图像基础。方法通过改变4D Extended Cardiac-Torso(XCAT)解剖结构输入参数,构建70组男性和女性仿真人体数据,并按照5∶1∶1比例划分为训练集、验证集和独立测试集。其中每组数据包括治疗前的双能CT(DECT)以及发生生理形变后的CBCT,后者反映放疗过程中的患者变化。使用迭代分解算法对双能CT进行物质分解,分别得到骨分解图(MDIB)和软组织分解图(MDIST)。构建基于断层图像的2D CycleGAN网络实现从CBCT到MDI的模态转化,并保留CBCT所代表的放疗当天真实解剖结构。网络以CBCT、MDIB和MDIST为输入,输出为治疗当日的MDIB与MDIST。为测试集患者构建与CBCT解剖结构相同的DECT并获得MDIB与MDIST作为真值,定量评估模型合成双能物质分解图的性能表现。结果在仅使用传统方案约13.8%的投影数量和辐射剂量情况下,测试集中的10套单能稀疏重建CBCT被模型成功转换成了与放疗当日解剖结构一致的MDIB与MDIST。合成的MDIB和MDIST与真值相比,结构相似性指数(SSIM)分别为0.983±0.006和0.988±0.005;均方根误差(RMSE)分别为0.017±0.005和0.019±0.004;峰值信噪比(PSNR)分别为35.515±2.081和34.409±1.510。模型训练耗时约18 h 51 min,合成每张MDI图像耗时约0.65 s。结论基于低剂量稀疏重建CBCT,本研究构建的2D CycleGAN网络可以实现跨模态、高保真的双能物质分解图像转化,有望在现有临床平台上为在线自适应放疗、离子放疗计划设计、剂量重建与监控等应用场景提供新型智能成像方法。

Objective To synthesize dual-energy material decomposition images(MDI)of anatomies consistent with the low-dose single-energy cone beam CT(CBCT)images on the treatment day,aiming to provide quantitative images for clinical application scenarios such as online adaptive radiotherapy(ART)and dose reconstruction.Methods Anthropomorphic data on 70 cases of males and females were generated by changing the anatomical input parameters of a 4D extended cardiac-torso(XCAT)phantom.These data were categorized into a training set,a validation set,and an independent test set at a ratio of 5∶1∶1.Each set consisted of pre-treatment dual-energy CT(DECT)images and post-physiological deformation CBCT images,which reflected changes in the patients during radiotherapy.An iterative decomposition algorithm was employed to conduct material decomposition of DECT images to derive material decomposition images of bone(MDIB)and material decomposition images of soft tissues(MDIST).A 2D CycleGAN network based on tomographic images was constructed to convert CBCT images to MDI while preserving the real anatomies on the treatment day reflected in CBCT images.CBCT images,MDIB,and MDIST were input into the network,which then output MDIB and MDIST on the treatment day.DECT images consistent with anatomies reflected in the CBCT images were constructed for the patients in the independent test set.As a result,MDIB and MDIST were synthesized,serving as the ground true images used to quantitatively evaluate the model performance in synthesizing dual-energy MDI.Results Under merely about 13.8%of the conventional projections and radiation dose,the model successfully converted 10 sets of single-energy CBCT images derived from sparse-view reconstruction in the test set into the MDIB and MDIST consistent with anatomies on the treatment day.Compared with the ground true images,the synthesized MDIB and MDIST showed structural similarity index(SSIM)values of 0.983±0.006 and 0.988±0.005,root-mean-square error(RMSE)values of 0.017±0.005 and 0.019±0.004,and peak signal-to-noise ratio(PSNR)values of 35.515±2.081 and 34.409±1.510,respectively.It took about 18 hours and 51 minutes to train the model and about 0.65 s to synthesize each MDI.Conclusions The 2D CycleGAN network developed in this study can synthesize cross-modal and high-fidelity dual-energy MDIs based on low-dose CBCT images derived from sparse-view reconstruction.Thus,using existing clinical platforms,it is expected to provide a novel smart imaging approach for clinical applications such as online adaptive radiotherapy,ion therapy planning,and dose reconstruction and monitoring.