[1]谢玉鑫,吴慧欣,吴亚平,等.基于三维Transformer去噪扩散模型的MRI到PET模态转换[J].中国医学物理学杂志,2026,43(9):1153-1161.[doi:DOI:10.3969/j.issn.1005-202X.2026.09.001]
 XIE Yuxin,WU Huixin,et al.MRI to PET modality translation using a 3D Transformer-based denoising diffusion model[J].Chinese Journal of Medical Physics,2026,43(9):1153-1161.[doi:DOI:10.3969/j.issn.1005-202X.2026.09.001]
点击复制

基于三维Transformer去噪扩散模型的MRI到PET模态转换()

《中国医学物理学杂志》[ISSN:1005-202X/CN:44-1351/R]

卷:
43卷
期数:
2026年第9期
页码:
1153-1161
栏目:
医学影像物理
出版日期:
2026-09-30

文章信息/Info

Title:
MRI to PET modality translation using a 3D Transformer-based denoising diffusion model
文章编号:
1005-202X(2026)09-1153-09
作者:
谢玉鑫1; 4; 吴慧欣2; 吴亚平3; 王梅云3; 4; 吕朋举4
1.华北水利水电大学信息工程学院, 河南 郑州 450046;2.河南省科学院空天信息研究所, 河南 郑州 450046;3.河南省人民医院医学影像科, 河南 郑州 450003;4.河南省科学院, 河南 郑州 450046
Author(s):
XIE Yuxin1;  4;  WU Huixin2;  WU Yaping3;  WANG Meiyun3;  4;  L?Pengju4
1. School of Information Engineering, North China University of Water Resources and Electric Power, Zhengzhou 450046, China 2. Institute of Aerospace Information, Henan Academy of Sciences, Zhengzhou 450046, China 3. Department of Medical Imaging, Henan Provincial Peoples Hospital, Zhengzhou 450003, China 4. Henan Academy of Sciences, Zhengzhou 450046, China
关键词:
磁共振成像; 正电子发射断层成像; 扩散模型; Transformer; 医学图像生成
Keywords:
magnetic resonance imaging positron emission tomography diffusion model Transformer medical image generation
分类号:
R318
DOI:
DOI:10.3969/j.issn.1005-202X.2026.09.001
文献标志码:
A
摘要:
目的:探索一种基于结构磁共振成像(MRI)合成功能性正电子发射断层成像(PET)的有效方法,旨在不依赖额外放射性示踪剂的情况下,实现对脑部代谢分布的可靠重建。方法:提出一种三维Transformer去噪扩散生成模型(3D T-DDPM),在条件扩散概率模型框架下完成MRI到PET的跨模态映射。模型以三维Swin Transformer作为主体网络,通过窗口自注意力机制有效建模三维体数据中的长距离空间依赖关系,并在浅层特征提取阶段引入卷积块注意力模块(CBAM),结合通道注意力与空间注意力对关键解剖结构和代谢相关区域进行自适应增强。结果:在私有大规模MRI-PET配对数据集上的实验结果表明,本研究提出的3D T-DDPM在合成PET图像的结构保真度、代谢分布一致性以及整体视觉质量方面均优于多种主流跨模态生成模型。与现有扩散模型和对抗生成模型相比,本研究能够更准确地恢复脑部解剖细节,并在功能相关区域呈现出更加稳定且合理的代谢表达。消融实验进一步验证三维Transformer主干和CBAM注意力模块的有效性,两者在提升长程结构建模能力和强化关键区域特征表达方面均发挥重要作用,共同促进跨模态映射精度的提升。结论:该方法为PET成像提供一种低经济、无辐射的替代方案,同时为无创脑功能成像及相关临床研究提供新的技术思路。
Abstract:
Abstract: Objective To explore an effective method for synthesizing functional positron emission tomography (PET) from structural magnetic resonance imaging (MRI), enabling reliable reconstruction of brain metabolic distribution without relying on additional radioactive tracers. Methods A three-dimensional Transformer-based denoising diffusion model (3D T-DDPM) was developed to perform cross-modal mapping from MRI to PET within the framework of a conditional diffusion probabilistic model. Taking the 3D Swin Transformer as the backbone network, this model effectively modeled long-range spatial dependencies within 3D volumetric data through a window-based self-attention mechanism. A convolutional block attention module was integrated into the shallow feature extraction stage to implement both channel and spatial attention for adaptively enhancing key anatomical structures and metabolically relevant regions. Results Experimental results on a private large-scale MRI-PET paired dataset showed that the proposed 3D T-DDPM outperformed several mainstream cross-modal generative models in terms of structural fidelity, metabolic distribution consistency, and overall visual quality of the synthesized PET images. Compared with existing diffusion models and adversarial generative models, this method could restore brain anatomical details more accurately and present more stable and reasonable metabolic patterns in functionally relevant regions. Ablation experiments further verified the effectiveness of the 3D Transformer backbone and convolutional block attention module. Both effectively enhanced long-range structural modeling and strengthened the feature representation of key regions, thereby improving the accuracy of cross-modal mapping. Conclusion The proposed method can serve as a low-cost, radiation-free alternative to PET imaging, and provide new technical insights for non-invasive brain function research and related clinical research.

相似文献/References:

[1]王水花,张煜东.压缩感知磁共振成像技术综述[J].中国医学物理学杂志,2015,32(02):158.[doi:10.3969/j.issn.1005-202X.2015.02.002]
[2]马海涵,游 箭,唐文国,等.成人小脑髓母细胞瘤MR成像和NSE表达水平[J].中国医学物理学杂志,2015,32(02):207.[doi:10.3969/j.issn.1005-202X.2015.02.012]
[3]武 杰,袁航英,严 峻,等.医用核磁共振成像设备的风险因素分析与管理[J].中国医学物理学杂志,2014,31(03):4918.[doi:10.3969/j.issn.1005-202X.2014.03.016]
[4]张煜东,王水花.磁共振成像加速方法[J].中国医学物理学杂志,2014,31(04):5015.[doi:10.3969/j.issn.1005-202X.2014.04.010]
[5]万俊,聂生东,王远军,等.基于MRI的脑肿瘤分割技术研究进展[J].中国医学物理学杂志,2013,30(04):4266.[doi:10.3969/j.issn.1005-202X.2013.04.013]
[6]曾雁冰,张战胜,辛学刚,等.反演法在磁共振射频场设计与优化中的应用[J].中国医学物理学杂志,2013,30(04):4272.[doi:10.3969/j.issn.1005-202X.2013.04.014]
[7]赵 琳,张军,张新宇,等.CT和MRI诊断脑神经胶质瘤52例分析[J].中国医学物理学杂志,2015,32(04):464.[doi:10.3969/j.issn.1005-202X.2015.04.003]
 [J].Chinese Journal of Medical Physics,2015,32(9):464.[doi:10.3969/j.issn.1005-202X.2015.04.003]
[8]赵 琳,张军,张新宇,等.CT和MRI诊断脑神经胶质瘤52例分析[J].中国医学物理学杂志,2015,32(04):464.[doi:10.3969/j.issn.1005-202X.2015.04.003]
[9]骆众星,石健强,谢斯栋,等.T1 FLAIR PROPELLER序列在3.0T磁共振颅脑增强成像上对抑制各种伪影的应用[J].中国医学物理学杂志,2015,32(06):878.[doi:doi:10.3969/j.issn.1005-202X.2015.06.025]
 [J].Chinese Journal of Medical Physics,2015,32(9):878.[doi:doi:10.3969/j.issn.1005-202X.2015.06.025]
[10]管春,朱庆永.参考制导的压缩感知快速磁共振重建方法[J].中国医学物理学杂志,2016,33(3):217.[doi:10.3969/j.issn.1005-202X.2016.03.001]

备注/Memo

备注/Memo:
【收稿日期】2026-01-17 【基金项目】国家自然科学基金(82441022, 82441023);河南省空天信息智能处理与集成应用工程研究中心开放课题(KTSOP202501) 【作者简介】谢玉鑫,硕士研究生,研究方向:人工智能与医学影像,E-mail: xyxyuxin2025@163.com 【通信作者】吴慧欣,博士,教授,研究方向:图形图像处理,E-mail: wuhuixin@ncwu.edu.cn;王梅云,主任医师,研究方向:医学影像学,E-mail: mywang@zzu.edu.cn(吴慧欣与王梅云为共同通信作者)
更新日期/Last Update: 2026-09-30