基于期望最大化算法的扩散弛豫定量方法研究  

Research on Diffusion-Relaxation Quantitative Analysis Method Based on Expectation-Maximization Algorithm

作　　者：高清宇 王磊[1,2] 王涛 王前锋 马超 王丽嘉 Qingyu Gao;Lei Wang;Tao Wang;Qianfeng Wang;Chao Ma;Lijia Wang(School of Health Science and Engineering,University of Shanghai for Science and Technology,Shanghai;Department of Radiology,Changhai Hospital of Shanghai,Naval Medical University,Shanghai;Department of Nuclear Medicine,Changhai Hospital of Shanghai,Naval Medical University,Shanghai;Institute of Science and Technology for Brain-Inspired Intelligence,Fudan University,Shanghai;School of Electronic and Information Engineering,Tongji University,Shanghai)

机构地区：[1]上海理工大学健康科学与工程学院,上海 [2]海军军医大学第一附属医院影像医学科,上海 [3]海军军医大学第一附属医院核医学科,上海 [4]复旦大学类脑智能科学与技术研究院,上海 [5]同济大学电子与信息工程学院,上海

出　　处：《建模与仿真》2025年第2期631-641,共11页Modeling and Simulation

摘　　要：目的:通过结合无监督机器学习算法和混合弛豫扩散磁共振成像(MRI),实现对胰腺癌的虚拟病理特征预测。方法:构建胰腺癌裸鼠模型,并使用11.7-T MRI扫描仪获取了包含多个b值(0、150、500、1500 s/mm^(2))和多个回波时间(25、50、75、100 ms)的扩散加权成像(DWI)数据。利用高斯混合模型和期望最大化算法,建立了一种混合弛豫扩散定量分析方法,用以计算肿瘤中上皮、间质和管腔成分的体积分数参数图,并与病理结果进行比较,以验证所建立分析方法的可行性。结果:超高场MRI下胰腺癌裸鼠模型弛豫扩散成像图像分辨率达到0.5×0.5×1 mm^(3),胰腺癌肿瘤组织的T2值范围为39至90 ms。基于混合扩散弛豫模型,我们建立的期望最大化算法能够快速计算出肿瘤组织中上皮、间质和管腔三种组织成分的相应权重。对于两种不同纤维化程度的肿瘤,基于混合扩散弛豫参数图计算出的肿瘤上皮(58.63%/86.55%)、间质(41.00%/13.18%)和管腔(1.65%/1.19%)成分的体积分数与病理结果相似。结论:本研究提出的无监督混合扩散弛豫定量分析方法为肿瘤组织内间质、上皮和管腔比例的虚拟病理预测提供了一种新手段。Objective:The aim of this study is to predict the virtual pathology of pancreatic cancer by integrating unsupervised machine learning algorithms with hybrid relaxation diffusion magnetic resonance imaging(MRI).Methods:We established a pancreatic tumor-bearing nude mouse model and acquired diffusion-weighted imaging data with multiple b-values(0,150,500,1500 s/mm^(2))and multiple echo times(25,50,75,100 ms)using an 11.7-T MRI scanner.Utilizing a Gaussian mixture model and the Expectation-Maximization algorithm,we developed a quantitative analysis method for hybrid relaxation diffusion to calculate the volume fraction parameter maps of epithelial,stromal,and lumen components in tumors,and then compared with the pathological results to verify the feasibility of the method.Results:Relaxation diffusion imaging of the pancreatic cancer-bearing nude mouse model performed under ultra-high field MRI was feasible,with an image resolution of 0.5×0.5×1 mm^(3),and the T2 values of tumor tissue ranged from 39 to 90 ms.Based on the assumptions of the hybrid diffusion relaxation model,our established Expectation-Maximization algorithm could rapidly calculate the corresponding weights of the epithelial,stromal,and lumen components in tumor.In two tumors with different degrees of fibrosis,the volumes of epithelial(58.63%/86.55%),stromal(41.00%/13.18%),and lumen(1.65%/1.19%)components calculated based on the hybrid diffusion relaxation parameter maps were similar to the pathological results.Conclusion:The unsupervised hybrid diffusion relaxation quantitative analysis method proposed in this study provides a new means for the virtual pathological prediction of the proportions of stroma,epithelium,and lumen within tumor.

关 键 词：超高场磁共振成像 混合扩散弛豫 无监督学习 期望最大化算法 胰腺癌 

分 类 号：R445.2[医药卫生—影像医学与核医学]