基于深度学习构建高频超声-病理图像配准及组织成分预测模型的体外研究  

In vitro study of constructing a high-frequency ultrasound-pathology image registration and tissue composition predictive model based on deep learning

作　　者：周文文 柴志菲 徐明[1] 黄通毅 赵泽[4] 任菲[4] 张晓儿 谢晓燕[1] ZHOU Wenwen;CHAI Zhifei;XU Ming;HUANG Tongyi;ZHAOZe;REN Fei;ZHANG Xiaoer;XIE Xiaoyan(Department of Medical Ultrasound,The First Affiliated Hospital,Sun Yat-sen University,Guangzhou 510080,Guangdong Province,China;Center for Medical Ultrasound,Suzhou Municipal Hospital,The Affiliated Suzhou Hospital of Nanjing Medical University,Suzhou 215002,Jiangsu Province,China;School of Computer Science,University of Chinese Academy of Sciences,Beijing 101408,China;Institute of Computing Technology,Chinese Academy of Sciences,Beijing 100190,China)

机构地区：[1]中山大学附属第一医院超声医学科,广东广州510080 [2]南京医科大学附属苏州医院苏州市立医院超声中心,江苏苏州215002 [3]中国科学院大学计算机学院,北京101408 [4]中国科学院计算技术研究所,北京100190

出　　处：《肿瘤影像学》2024年第6期586-592,共7页Oncoradiology

基　　金：国家自然科学基金重大项目(92059201);国家自然科学基金面上项目(82071951);国家自然科学基金青年项目(82402297)。

摘　　要：目的:探讨基于深度学习(deep learning,DL)网络,在体外模型中建立高频超声(high-frequency ultrasound,HFUS)-病理图像融合配准方法并构建病理组织成分模型的可行性及预测价值。方法:制备包含4种不同生物组织和定位颗粒的60个体外仿肿瘤模型,在物理匹配下获取同一切面的HFUS图像和全视野切片图像(wide slide image,WSI)。对图像进行质量控制和筛选,在WSI中沿组织边缘手动勾画出感兴趣区域(region of interest,ROI)并转移至HFUS图像上。以原始图像及对应ROI为数据集,按照13∶1∶1分为训练集(n=462)、验证集(n=34)和测试集(n=38)。通过迁移学习DeepLabV3、FCN-50和MobileNetV3网络构建DL模型,并输出自动分割图像,最终采用像素准确度(pixel accuracy,PA)、精确度、灵敏度和F1-score指标量化和比较训练集和测试集中各模型预测效能。结果:基于DeepLabV3、FCN-50和MobileNetV3网络的DL模型自动分割测试集中不同组织成分的准确度和相似度均较高,其中MobileNetV3模型效能最优,其PA为91.4%、F1-score为87.1%。不同网络模型效能之间差异无统计学意义(P>0.05)。各模型预测体外不同组织成分的效能差异之间差异有统计学意义(P<0.001),其中肝脏组织效果最佳。结论:本研究构建的超声-病理融合模型能够有效地识别超声图像中体外组织成分,为进一步临床应用提供方法学依据。Objective:To investigate the feasibility of fusion registration of high-frequency ultrasound(HFUS)-pathology images and predictive value of model for predicting the pathological tissue components in in vitro models based on deep learning(DL)networks.Methods:Sixty in vitro mimetic tumor models containing four different biological tissues and localized particles were prepared.HFUS images and wide slide images(WSI)of the same slide were obtained under physical registration.The obtained images were quality controlled and selected,and the region of interest(ROI)was manually outlined along the edges of the tissues in the WSI and then transferred to HFUS images.The datasets were consisted of original images and corresponding ROIs and were divided into the training set(n=462),validation set(n=34)and testing set(n=38)at the ratio of 13∶1∶1.DL models were developed via transfer learning DeepLabV3,FCN-50 and MobileNetV3 networks.The pixel accuracy(PA),precision,recall and F1-score were used to quantify and compare the performance of each model in the training and testing datasets.The automatically segmented images were output.Results:The DL models based on DeepLabV3,FCN-50 and MobileNetV3 networks had high accuracy and similarity for automatically segmentation of different tissue components in the testing set, and the MobileNetV3 model outperformed others withthe PA of 91.4% and F1-score of 87.1%. There was no significant difference between performance of models (all P> 0.05). There werestatistically significant differences between the efficiencies of models for predicting different in vitro biological tissue components (allP<0.001), with the best of liver tissue. Conclusion: The constructed ultrasound-pathology fusion models in this study can effectivelyrecognize the in vitro tissue components in ultrasound images and provide the methodological basis for further clinical applications.

关 键 词：高频超声 体外模型 深度学习 全视野数字切片 图像配准 预测 

分 类 号：R445.1[医药卫生—影像医学与核医学]