Objective: To investigate the value of radiomics models based on magnetic resonance imaging (MRI) diffusion weighted imaging (DWI) and apparent diffusion coefficient (ADC) maps in distinguishing benign and malignant thyroid nodules. Methods: A cross-sectional study. Clinical data of 148 thyroid nodules (50 benign, 98 malignant) from 140 patients who underwent thyroid MRI examination in Cancer Hospital & Shenzhen Hospital, Chinese Academy of Medical Sciences between January 2019 and December 2022 were retrospectively analyzed. The nodules were used as the study units, and a leave-one-out method was used to randomly divide the nodules into a training set and a test set at a 7∶3 ratio. Region of interest was segmented and radiomics features were extracted from the DWI and ADC images. In the training set, feature selection was performed using inter-observer agreement analysis, U-test, least absolute shrinkage and selection operator algorithm, and correlation analysis. Four classifiers, including support vector machine (SVM), random forest (RF), k-nearest neighbors (KNN) and logistic regression (LR) were used to build models with the selected features, including the DWI models, ADC models, and combined models. The models were independently tested in the test set. The performance of the radiomics models in distinguishing benign and malignant thyroid nodules was evaluated using the receiver operating characteristic (ROC) curve, with pathological results as the gold standard. Results: Of the 140 patients, there were 40 males and 100 females, with a mean age of (38.4±12.2) years. After feature selection, 11 DWI features and 11 ADC features were used to build the models. In the training set, the AUC values of the combined models were higher than those of the corresponding DWI and ADC models. In the test set, the SVM combined model showed the best predictive performance, with an AUC of 0.873 (95%CI:0.740-0.954), accuracy of 75.6%, sensitivity of 46.7%, specificity of 90.0%, positive predictive value (PPV) of 70.0% and negative predictive value (NPV) of 77.1%, while the RF combined model had an AUC of 0.836 (95%CI:0.695-0.929), accuracy of 77.8%, sensitivity of 40.0%, specificity of 96.7%, PPV of 85.7% and NPV of 76.3%, the KNN combined model had an AUC of 0.832 (95%CI:0.691-0.927), accuracy of 77.8%, sensitivity of 33.3%, specificity of 100%, PPV of 100% and NPV of 75.0%, the LR combined model had an AUC of 0.813 (95%CI:0.669-0.914), accuracy of 77.8%, sensitivity of 60.0%, specificity of 86.7%, PPV of 69.2% and NPV of 81.3%. Conclusions: Radiomics models based on DWI and ADC image features can effectively distinguish benign and malignant thyroid nodules. The SVM combined model had the best prediction performance.
目的： 探讨基于MRI扩散加权成像（DWI）和表观扩散系数（ADC）图像的影像组学模型鉴别诊断甲状腺结节良恶性的价值。 方法： 横断面研究。回顾性分析2019年1月至2022年12月中国医学科学院肿瘤医院深圳医院行甲状腺MRI检查的140例患者的148个甲状腺结节（良性50个，恶性98个）的临床资料。以结节为研究单位，使用留出法将甲状腺结节按照7∶3的比例随机分成训练集和测试集。对DWI和ADC图像进行感兴趣区勾画及组学特征提取，在训练集中采用观察者间一致性分析、U检验、最小绝对收缩和选择算子算法、相关性分析进行特征筛选，使用支持向量机（SVM）、随机森林（RF）、K最近邻（KNN）和逻辑回归（LR）4个分类器对选取的特征构建模型，包括DWI模型、ADC模型和联合模型，并在测试集中对模型进行测试。以甲状腺结节病理结果为金标准，应用受试者工作特征（ROC）曲线评价影像组学模型鉴别诊断甲状腺结节良恶性的效能。 结果： 本研究140例患者中，男40例，女100例，年龄（38.4±12.2）岁。经过特征筛选，11个DWI特征和11个ADC特征被用于构建模型。训练集中，基于同一分类器构建的不同模型间比较，联合模型的ROC曲线下面积（AUC）均高于相应的DWI模型和ADC模型。测试集中，SVM联合模型表现出最佳的模型预测效能［AUC为0.873（95%CI：0.740~0.954），准确度为75.6%，灵敏度为46.7%，特异度为90.0%，阳性预测值（PPV）为70.0%，阴性预测值（NPV）为77.1%］，其AUC高于RF联合模型［AUC为0.836（95%CI：0.695~0.929），准确度为77.8%，灵敏度为40.0%，特异度为96.7%，PPV为85.7%，NPV为76.3%］、KNN联合模型［AUC为0.832（95%CI：0.691~0.927），准确度为77.8%，灵敏度为33.3%，特异度为100%，PPV为100%，NPV为75.0%］以及LR联合模型［AUC为0.813（95%CI：0.669~0.914），准确度为77.8%，灵敏度为60.0%，特异度为86.7%，PPV为69.2%，NPV为81.3%］。 结论： 基于DWI和ADC图像特征的影像组学模型有助于鉴别诊断甲状腺结节良恶性，SVM联合模型的预测效能最佳。.