Page 74 - 南京医科大学自然版
P. 74

南京医科大学学报(自然科学版)                                  第44卷第7期
               ·958 ·                     Journal of Nanjing Medical University(Natural Sciences)   2024年7月


             ·临床研究·

              基于光谱CT和影像组学特征的机器学习模型和列线图预测甲

              状腺结节良恶性



              王 皓 ,杜昱升 ,王 友 ,白卓杰 ,杨 帆 ,何 健                       1,3*
                                                        2
                     1,2
                               2
                                                2
                                       2
               南京医科大学鼓楼临床医学院,江苏              南京   210008;南京医科大学第四附属医院放射科,江苏                南京   210031;南京大学
              1                                         2                                              3
              医学院附属鼓楼医院核医学科,江苏              南京 210008
             [摘    要] 目的:构建光谱 CT 参数和影像组学机器学习模型预测甲状腺结节良恶性。方法:回顾性分析行光谱 CT 增强扫
              描的 118 例甲状腺结节患者(143 个结节,其中包括 46 例良性结节和 97 例恶性结节)影像及临床资料,7∶3 随机分为训练集
             (n=100)和验证集(n=43)。采用差异性检验、组间一致性评估以及最小绝对收缩和选择算子算法(least absolute shrinkage and
              selection operator,LASSO)筛选特征并计算影像组学评分。运用决策树(decision tree,DT)、随机森林(random forest,RF)、极
              端梯度提升树(extreme gradient boosting,XGBoost)、支持向量机(support vector machine,SVM)、K 最近邻(K⁃nearest neighbor,
              KNN)和逻辑回归(logistic regression,LR)6种机器学习算法进行建模,筛选最佳的模型并构建列线图。结果:XGBoost 模型在
              验证集中性能最好(曲线下面积:0.938;准确度:86.05%;灵敏度:89.29%;特异度:80.00%),标准化碘值、影像组学评分与年龄
              是重要且有效的预测因素,构建的列线图具有良好的性能。结论:结合光谱CT和影像组学的机器学习模型及列线图能够为甲
              状腺结节良恶性的非侵入性预测提供高准确性的参考。
             [关键词] 影像组学;机器学习;光谱CT;甲状腺结节
             [中图分类号] R814.42                    [文献标志码] A                     [文章编号] 1007⁃4368(2024)07⁃958⁃08
              doi:10.7655/NYDXBNSN240049


              Machine learning models and nomograms based on spectral CT and radiomic features for
              predicting the benignancy and malignancy of thyroid nodules

                        1,2          2           2          2          2        1,3*
              WANG Hao ,DU Yusheng ,WANG You ,BAI Zhuojie ,YANG Fan ,HE Jian
                                                                                        2
              1 Drum Tower Clinical Medical College of Nanjing Medical University,Nanjing 210008;Department of Radiology,
              the Fourth Affiliated Hospital of Nanjing Medical University,Nanjing 210031;Department of Nuclear Medicine,
                                                                                  3
              Gulou Hospital Affiliated to Nanjing University School of Medicine,Nanjing 210008,China


             [Abstract] Objective:To establish the applicability of a predictive model based on spectral computed tomography(CT)parameters
              and radiomics features through machine learning for differentiating benign and malignant thyroid nodules. Methods:A retrospective
              analysis was conducted on imaging and clinical data from 118 patients with thyroid nodules who underwent spectral CT enhancement
              scanning(143 nodules,comprise of 46 benign and 97 malignant nodules). These patients were randomly divided into a training set
             (n=100)and a validation set(n=43)in a 7∶3 ratio. Discriminative testing,intraclass correlation coefficient(ICC),and the least
              absolute shrinkage and selection operator(LASSO)were employed to select features and calculate a radiomics score(Radscore). Six
              machine learning algorithms including decision tree(DT),random forest(RF),extreme gradient boosting(XGBoost),support vector
              machine(SVM),K⁃nearest neighbors(KNN),and logistic regression(LR)were utilized to develop models. The optimal model was
              selected to construct nomograms. Results:The XGBoost model was demonstrated to be the best in the validation set(AUC:0.938;
              accuracy:86.05%;sensitivity:89.29%;specificity:80.00%),with normalized iodine concentration(NIC),Radscore,and age identified
              as significant predictive factors. The ensuing nomograms exhibited robust performance. Conclusion:The machine learning model that
              combines spectral CT and radiomics features with the nomograms provides a highly accurate reference for non⁃invasive prediction of

             [基金项目] 江苏省青年医学重点人才(QNRC2016041);南京医科大学科技发展基金一般项目(NMUB20230037)
              ∗
              通信作者(Corresponding author),E⁃mail:hjxueren@126.com
   69   70   71   72   73   74   75   76   77   78   79