摘要
目的:探讨基于腹部平扫CT人工智能(AI)模型评估肝脂肪含量(LFC)的价值。方法:回顾性将2022年7月-2023年5月在本院接受腹部CT平扫及定量CT(QCT)检查的840例体检者纳入本研究,按照8∶2的比例将所有受试者分为训练集(n=672)及测试集(n=168)。在肝左叶、右前叶和右后叶的外周区域各放置1个ROI,采用QCT测量,以3个ROI的脂肪分数(FF)的平均值作为LFC。依据LFC测量结果,将受试者分为正常肝脏组(FF为<5%)、轻度脂肪肝组(FF为5%~14%)、中度脂肪肝组(FF为14%~28%)和重度脂肪肝组(FF为>28%)。在腹部平扫CT图像上沿肝脏轮廓勾画ROI,将连续3层图像上勾画的肝脏ROI作为2D-ROI,全肝图像上勾画的ROI作为3D-ROI。将临床数据和肝脏影像学定量参数(包括肝脏CT值、体积、大小等)作为机器学习(ML)特征,采用8种ML算法构建临床-影像ML模型。采用Python开源工具包Pyradiomics(v3.0.1)分别提取2D-ROI和3D-ROI的影像组学(Rad)特征,再用相同的8种ML算法基于2D-Rad和3D-Rad的组学特征分别构建肝脏脂肪变四分类诊断模型。分别采用Pearson相关分析及Bland-Altman分析观察各模型测量的FF值与QCT测量的FF值的一致性,采用受试者工作特征(ROC)曲线评估各模型分类诊断肝脂肪变的效能。结果:在基于临床-影像、2D-Rad和3D-Rad构建的各种ML模型中,基于Bagging决策树算法的2D-Rad模型(2D-Rad Baggiing)诊断肝脂肪变的AUC(0.982)最大;2D-Rad DT和2D-Rad GP分类模型的敏感度(0.861)最高,2D-Rad Bagging和2D-Rad DT分类模型的特异度(0.950)最高,2D-Rad Bagging、2D-Rad DT和2D-Rad GP模型的符合率(0.857)最高,2D-Rad GP模型的精确率和F1得分最高。各模型测量的FF值与QCT测量的FF的一致性良好(绝大多数差值位于均值±1.96标准差的范围内),且呈高度正相关(r=0.920~0.990)。结论:基于腹部平扫CT的AI模型分类诊断肝脂肪变的效能较高,且所测LFC与QCT结果的一致性良好。
Objective:To explore the effectiveness of AI models in estimating liver fat content based on non-contrast abdominal CT images and to compare with QCT.Methods:Retrospectively,840 physical examinees who underwent non-contrast abdominal CT and quantitative CT(QCT)examinations in our hospital from July 2022 to May 2023 were included.All subjects were divided into a trai-ning set(n=672)and a test set(n=168)at an 8∶2 ratio.One region of interest(ROI)was placed in the peripheral area of the left hepatic lobe,right anterior lobe,and right posterior lobe,respectively.QCT was used for measurement,and the mean value of fat fractions(FFs)from the three ROIs was defined as liver fat content(LFC).According to LFC measurements,subjects were classified into four groups:normal liver(FF<5%),mild fatty liver(FF in 5%~14%),moderate fatty liver(FF in 14%~28%),and severe fatty liver(FF>28%).On plain abdominal CT images,ROIs were contoured along the liver outline.The liver ROIs contoured on three consecutive slices were defined as 2D-ROIs(two-dimensional regions of interest),while those contoured on the entire liver image were defined as 3D-ROIs(three-dimensional regions of interest).Clinical data and quantitative liver imaging parameters(including liver CT value,volume,size,etc.)were used as machine learning(ML)features,and eight ML algorithms were applied to construct clinical-imaging ML models.The open-source Python toolkit Pyradiomics(v3.0.1)was used to extract radiomic(Rad)features from 2D-ROIs and 3D-ROIs,respectively.Eight identical ML algorithms were then applied to develop 2D-Rad and 3D-Rad models for four-class diagnosis of hepatic steatosis based on Rad features.To observe the consistency of fat fraction(FF)values measured by each model and those measured by QCT,Pearson correlation analysis and Bland-Altman analysis were respectively used.The receiver operating characteristic(ROC)curve was adopted to evaluate the classification and diagnostic efficacy of each model for hepa-tic steatosis.Results:Among various ML models constructed based on clinical-imaging,2D-Rad,and 3D-Rad features,the 2D-Rad model based on the Bagging decision tree algorithm(2D-Rad Bagging)showed the largest area under the ROC curve(AUC=0.982).The 2D-Rad DT and 2D-Rad GP classification models had the highest sensitivity(0.861),while the 2D-Rad Bagging and 2D-Rad DT classification models had the highest specificity(0.950).The 2D-Rad Bagging,2D-Rad DT and 2D-Rad GP models showed the highest accuracy(0.857),and the 2D-Rad GP model had the highest precision rate and F1-score.The consistency between the FF values measured by each model and those measured by QCT was good(most differences were within the range of mean±1.96SD)and exhibited high positive correlation(r=0.920~0.990).Conclusion:The AI model based on non-contrast abdominal CT images demonstrates high efficacy in classifying and diagnosing hepatic steatosis,and the measured LFC shows good consistency with QCT results.
作者
张浩然
白臻
苏丹阳
刘金龙
马渊博
苗秋菊
杨慎宇
任向阳
杨晓鹏
ZHANG Hao-ran;BAI Zhen;SU Dan-yang(Department of Radiology,the First Affiliated Hospital of Zhengzhou University,Zhengzhou 450052,China)
出处
《放射学实践》
北大核心
2025年第8期1011-1017,共7页
Radiologic Practice
基金
2024年度郑州市基础研究与应用基础研究(2024ZZJCYJ047)
国家资助博士后研究人员计划(GZC20241549)。
关键词
肝脏病变
脂肪变性
脂肪含量
定量CT
人工智能
Hepatic lesion
Steatosis
Fat content
Quantitative computed tomography
Artificial intelligence
作者简介
张浩然(2000-),男,河南郑州人,硕士研究生,主要从事研影像技术工作。;通讯作者:杨晓鹏,E-mail:13837141925@163.com。
