基于三维深度分离网络的PET双示踪剂混合图像分离方法  

Dual-tracer PET image separation using three-dimensional depthwise separable convolution network

作　　者：唐大洋 胡德斌 齐宏亮 孙浩 韩彦江[3] 李翰威 张新明 潘智林 喻文杰 路利军[1] 陈宏文 TANG Dayang;HU Debin;QI Hongliang;SUN Hao;HAN Yanjiang;LI Hanwei;ZHANG Xinming;PAN Zhilin;YU Wenjie;LU Lijun;CHEN Hongwen(School of Biomedical Engineering,Southern Medical University,Guangzhou 510515,China;Department of Clinical Engineering,Nanfang Hospital,Southern Medical University,Guangzhou 510515,China;PET Center,Nanfang Hospital,Southern Medical University,Guangzhou 510515,China)

机构地区：[1]南方医科大学生物医学工程学院,广东广州510515 [2]南方医科大学南方医院医学工程科,广东广州510515 [3]南方医科大学南方医院PET中心,广东广州510515

出　　处：《中国医学物理学杂志》2025年第2期160-166,共7页Chinese Journal of Medical Physics

基　　金：国家重点研发计划(2023YFC2414601);广东省医学会医学工程学分会青年委员会基金(2022-GDMAYB-05);南方医科大学南方医院院长基金(2022B016)。

摘　　要：目的:提出一种基于三维深度分离网络方法用于^(18)F-FDG和^(18)F-FAPIPET双示踪剂混合图像分离成像。方法:收集120例同一患者在不同时间单独扫描的^(18)F-FDG和^(18)F-FAPIPET图像,本研究采用模拟的形式生成PET双示踪剂混合图像,首先对同一患者两种PET示踪剂图像进行配准保证空间位置匹配,然后对配准的PET图像进行前向投影生成弦图数据,将两种弦图数据累加得到混合弦图数据,随后采用最大似然期望法重建得到PET双示踪剂混合图像,输入到基于3DDSN架构的网络进行分离成像,从而得到两种单示踪剂的PET图像。结果:本文提出的方法相较于3DCNN方法,分离得到的^(18)F-FDG图像与真实^(18)F-FDG图像的结构相似性指数(SSIM)提升0.87%,峰值信噪比(PSNR)提升11.8%,归一化均方根误差(NRMSE)减小52%。分离得到的^(18)F-FAPI图像与真实^(18)F-FAPI图像的SSIM提升1.1%,PSNR提升17.0%,NRMSE减小51%。结论:本文方法可以很好地应用在PET双示踪剂同时成像上,减少患者的扫描次数、时间和金钱成本,为临床医生提供更精准和更丰富的诊断信息。Objective To propose a novel method based on three-dimensional depthwise separable convolution network(3D DSN)for the separation of PET images with dual tracers of ^(18)F-FDG and ^(18)F-FAPI.Methods A total of 120 pairs of ^(18)F-FDG and ^(18)F-FAPI PET images of the same patient scanned separately at different time points were collected,and the dual-tracer PET image was generated through simulation.After the image registration of PET images of two tracers for ensuring spatial position matching,the registered PET images were forward-projected to generate sinogram data,and the sinogram data of two tracers were accumulated to obtain mixed sinogram data.Subsequently,the dual-tracer PET image was reconstructed using maximum likelihood expectation maximization and input into a 3D DSN based network for image separation,thereby obtaining PET images of two single tracers.Results Compared with 3D CNN method,the proposed method increased the structure similarity index measure(SSIM)of the separated ^(18)F-FDG images to the real ^(18)F-FDG images by 0.87%,increased the peak signal-to-noise ratio(PSNR)by 11.8%,and reduced the normalized root mean square error(NRMSE)by 52%.The SSIM of the separated ^(18)F-FAPI images to the real ^(18)F-FAPI images increased by 1.1%,PSNR increased by 17.0%,and NRMSE decreased by 51%.Conclusion The proposed method can be effectively applied to simultaneous PET imaging with dual PET tracers,reducing the number of scans and costs in time and money,and providing clinical doctors more accurate and abundant diagnostic information.

关 键 词：正电子发射断层成像 双示踪剂成像 图像配准 深度分离网络 深度学习 

分 类 号：R318[医药卫生—生物医学工程] R817[医药卫生—基础医学]