基于CT图像的深度学习模型鉴别良恶性肾肿瘤  

CT-based deep learning model for differentiating renal cell carcinoma from benign renal tumor

作　　者：陈美容[1] 周涛 陈炫幸 钟向阳[1] 陈智慧 冯宝[3] 陈业航[3] 徐坤财 CHENMei-rong;ZHOU Tao;CHEN Xuan-xing(Department of Radiology,Jiangmen Central Hospital of Guangdong Province,Guangdong 529000,China)

机构地区：[1]江门市中心医院放射科,广东江门529000 [2]阳江市人民医院放射科,广东阳江529500 [3]桂林航天工业学院,广西桂林541000

出　　处：《放射学实践》2023年第4期468-473,共6页Radiologic Practice

摘　　要：目的:探讨基于CT图像的深度学习模型对肾脏良、恶性肿瘤的鉴别诊断价值。方法:回顾性搜集2008-2020年经病理证实且符合本研究要求的798位患者(共805例肾肿瘤)的临床和三期(平扫、皮髓质期和实质期)腹部CT影像资料。其中,来自本院的418例肾癌和78例肾良性肿瘤的资料用于建立影像组学模型和内部验证,来自另外两个研究中心和一个公共数据库(癌症医学图像数据库TCIA)的262例肾癌和47例肾良性肿瘤的资料作为独立外部验证集。使用ITK-SNAP3.6.0软件,在三期CT图像中选择肿瘤边缘显示较清楚的一期图像,选取肿瘤最大层面及其上、下相邻层面,沿病灶边缘手动勾画ROI,再通过软件的空间调整技术,使另外两期CT图像上肿瘤的边缘与勾画的ROI的边缘最大程度地拟合。使用ResNet50网络中的卷积核作为特征提取器,分别提取3期图像上肾肿瘤的影像组学特征,并采用Mann-Whitney U检验进行特征的筛选。对于同一肾肿瘤,分别建立基于单期和3期CT图像的深度学习模型,并对各模型的预测效能进行外部验证。随后,将训练集中良、恶性肿瘤组的样本数按3种比例(1∶1、1∶2、1∶3)进行设置,分别用于极限学习机(ELM)分类模型的训练,建立基于多期CT图像的3种深度学习模型,并对模型进行外部验证。采用AUC曲线评估深度学习模型对良恶性肾肿瘤的鉴别诊断效能,采用综合判别改善指数(IDI)评估模型预测能力的改善情况。结果:基于多期CT图像的预测模型预测恶性肿瘤的AUC(0.84)大于基于单期(平扫、皮髓质期和实质期)图像的3个AUC(0.78、0.79、0.77)。良性与恶性肿瘤的样本数比例分别为1∶1、1∶2和1∶3时,基于多期图像的预测模型的AUC分别为0.85、0.84和0.86。基于多期图像的预测模型与基于单期图像的3个预测模型比较,IDI值分别为0.1215、0.1209和0.0094(P均>0.05)。结论:基于多期CT图像的深度学习�Objective:To investigate the value of a CT-based deep learning model for the diffe-rentiation of renal carcinoma and benign renal tumor.Methods:Clinical and three-phase[(precontrast phase,PCP),(cortical medullary phase,CMP),(nephrographic phase,NP)]abdominal CT imaging data of 798 patients(a total of 805 renal tumors)with pathologically proven RCCs and benign renal tumors between 2008 and 2020 were retrospectively collected,including 418 RCCs and 78 benign tumors from our institution,as the training dataset for model development and internal validation.Patients from two independent institutions and a public database(the Cancer Imaging Archive,TCIA)were included as the external dataset for individual testing,including 262 RCCs and 47 benign tumors.Among the three phases,the phase in which the margin of the tumor was clearly showed was selected,and then on the slice of the tumor with maximum diameter the its upper and lower slices,regions of interest(ROIs)in the tumor was delineated using ITK-SNAP(Version 3.6.0)software,and then space adjustment technique of the software was used to maximize the fit of the CT images from other two phases with the delineated tumor region.The convolution kernel in ResNet50 network was used as the feature extractor to extract the features from the three phases CT images of renal tumor.The Mann-Whitney U-test was used for feature selection.Deep learning models based on single-phase and multi-phases CT images were built for the same renal tumor,respectively,and were validated with external data.Then,the sample numbers of benign and malignant tumor groups were set according to three ratios(1∶1,1∶2,1∶3)respectively,which were used for the training of extreme learning machine(ELM)classification model.The three deep learning models based on three phases CT images were established,and the prediction models were externally validated.The AUC curve was used to evaluate the differential diagnosis efficiency of the deep learning model for benign and malignant renal tumors,and the Integrated

关 键 词：肾肿瘤 深度学习 影像组学 体层摄影术 X线计算机 

分 类 号：R814.42[医药卫生—影像医学与核医学] R737.11[医药卫生—放射医学]