摘要
目的:基于术前^(18)F-氟代脱氧葡萄糖正电子发射断层扫描/计算机断层扫描(^(18)F-FDG PET/CT),构建并验证结合深度学习(DL)特征、手工影像组学(HCR)特征和临床参数的多模态模型,用于术前预测结直肠腺癌淋巴结转移(LNM)及评估预后价值。方法:回顾性分析2011年7月-2021年12月行术前^(18)F-FDG PET/CT检查并确诊为Ⅱ/Ⅲ期结直肠腺癌患者338例。按照8∶2随机将患者划分为训练集(n=270)和测试集(n=68)。使用ResNet50模型作为DL特征提取的基础模型,分别于PET及CT图像提取HCR特征及DL影像组学特征,采用Pearson相关性系数、最大相关最小冗余(mRMR)、最小绝对收缩与选择算子(LASSO)算法等筛选最优影像组学特征,采用单因素与多因素Logistic回归分析筛选临床参数中独立危险因素。分别基于HCR特征、DL特征、临床参数及三者组合特征,构建Logistic回归(LR)、支持向量机(SVM)、k近邻(KNN)、随机森林(RF)、极度随机树(ET)、极端梯度提升(XGBoost)、轻量级梯度提升机(LightGBM)及多层感知器(MLP)8种机器学习模型。绘制受试者操作特征(ROC)曲线及校准曲线,比较曲线下面积(AUC)、准确率、敏感度、特异度以评估模型的性能。结果:单因素及多因素Logistic回归分析显示临床参数中淋巴结最大短径(OR=2.645)及淋巴结SUVmax(OR=4.093)为预测LNM的独立风险因素(P<0.05)。通过整合DL特征、HCR特征及临床参数构建的联合模型表现出优秀的预测性能,其在训练集与测试集中的AUC分别为0.829(95%CI:0.780~0.878)和0.735(95%CI:0.614~0.857),优于单一临床模型,校准曲线表明模型具有良好的预测一致性。Kaplan-Meier生存曲线分析显示LNM显著增加患者复发或转移风险。结论:基于术前^(18)F-FDG PET/CT深度学习的影像组学模型能够较准确预测结直肠腺癌患者的LNM,为术前精准分期及个体化治疗提供重要依据。
Objective:The aim of this study was to create and authenticate a combined model based on preoperative ^(18)F-fluorodeoxyglucose positron emission tomography/computed tomography(^(18)F-FDG PET/CT)integrated with deep learning(DL)features,handcrafted radiomics(HCR),and clinical parameters for predicting lymph node metastasis(LNM)in colorectal adenocarcinoma and to assess its impact on prognosis.Methods:A retrospective analysis was conducted on 338 patients with stage Ⅱ/Ⅲ colorectal adenocarcinoma who underwent preoperative ^(18)F-FDG PET/CT from July 2011 to December 2021.The cohort was randomly divided into training(n=270)and test(n=68)sets at an 8:2 ratio.The ResNet50 model was used as the backbone for DL feature extraction.HCR and DL radiomic features were extracted from PET and CT images and optimal features were selected using Pearson correlation coefficient,maximum correlation minimum redundancy(mRMR),and least absolute shrinkage and selection(LASSO).Univariate and multivariate logistic regression identified independent clinical risk factors.Predictive models were constructed based on HCR,DL,clinical parameters,and their combinations using eight machine learning algorithms,including logistic regression(LR),support vector machine(SVM),k-nearest neighbors(KNN),random forest(RF),extremely randomized trees(ET),eXtreme Gradient Boosting(XGBoost),Light Gradient Boosting Machine(LightGBM),and multilayer perceptron(MLP).Model performance was evaluated using receiver operating characteristic(ROC)curves,calibration curves,area under the curve(AUC),accuracy,sensitivity,and specificity.Results:The results of both univariate and multivariate logistic regression analysis indicate that maximum short-axis diameter of lymph nodes(OR=2.645)and lymph nodes SUVmax(OR=4.093)as independent risk factors for predicting(P<0.05).The combined model integrating DL features,HCR features,and clinical parameters demonstrated superior performance,achieving AUCs of 0.829(95%CI:0.780~0.878)and 0.735(95%CI:0.614~0.857)in the training and test sets,respectively,outperforming clinical models alone.Calibration curves indicated strong predictive consistency.Kaplan-Meier survival curves revealed that LNM significantly increased the risk of recurrence or metastasis.Conclusion:Deep learning models based on preoperative ^(18)F-FDG PET/CT radiomic features accurately predict LNM in patients with colorectal adenocarcinoma,providing critical insights for precise preoperative staging and personalized treatment strategies.
作者
王冰
屈伟
郑龙
胡添源
张占文
WANG Bing;QU Wei;ZHENG Long(Department of Nuclear Medicine,the Second Affiliated Hospital of Xi'an Jiaotong University,Xi'an 710004,China)
出处
《放射学实践》
北大核心
2026年第3期307-316,共10页
Radiologic Practice
基金
广东省自然科学基金-面上项目基金资助项目(2023A1515011300)。
关键词
正电子发射断层扫描
影像组学
深度学习
结直肠腺癌
淋巴结转移
Positron emission tomography
Radiomics
Deep learning
Colorectal adenocarcinoma
Lymph node metastasis
