Zhang Yutong, Jiang Jue, Chen Aqian, Zhang Dong, Wang Lirong, Yuan Xin, He Xin, Yu Shanshan, Wang Juan, Zhou Qi
Department of Ultrasound, The Second Affiliated Hospital of Xi 'an Jiaotong University, Xi'an, China.
National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, National Engineering Research Center for Visual Information and Applications, and Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University, Xi'an, China.
Front Endocrinol (Lausanne). 2025 Aug 6;16:1635122. doi: 10.3389/fendo.2025.1635122. eCollection 2025.
Partially cystic thyroid nodules (PCTNs) with malignant potential are frequently underestimated due to limited recognition of their sonographic characteristics.
This retrospective analysis included 486 PCTNs identified between March 2021 and September 2022. Machine learning (ML) was employed to quantitatively evaluate the overall ultrasound characteristics of the whole nodule as well as the internal ultrasound characteristics of its solid part. Three diagnostic models were constructed based on different sets of ultrasound data. The dataset was split into training and testing subsets at a 7:3 ratio. Key ultrasound characteristics such as marked hypoechogenicity, calcifications, solid component≥50%, and unclear internal margins were emphasized.
Among the models, the integrated one- incorporating both overall-nodule and internal solid-part characteristics-achieved superior diagnostic performance, with an area under the curve (AUC) of 0.96 (0.93-0.99) on the test data. The model demonstrated an accuracy of 0.91 (0.85-0.95), a sensitivity of 0.88 (0.73-0.97), a specificity of 0.92 (0.85-0.96), a negative predictive value of 0.96 (0.91-0.99), and a positive predictive value of 0.77 (0.61-0.89). This comprehensive model significantly outperformed the model utilizing only overall nodule characteristics (AUC = 0.85, P = 2.35e-6), and demonstrated comparable effectiveness to the model based solely on internal characteristics (AUC = 0.93, P = 1.01e-1).
The results support the clinical utility of an ML-driven approach that integrates comprehensive ultrasound metrics for the reliable identification of malignant PCTNs.
具有恶性潜能的部分囊性甲状腺结节（PCTN）因其超声特征的识别有限而常常被低估。
这项回顾性分析纳入了2021年3月至2022年9月期间识别出的486个PCTN。采用机器学习（ML）定量评估整个结节的总体超声特征及其实性部分的内部超声特征。基于不同的超声数据集构建了三种诊断模型。数据集按7:3的比例分为训练子集和测试子集。强调了关键超声特征，如显著低回声、钙化、实性成分≥50%以及内部边界不清。
在这些模型中，整合了总体结节和内部实性部分特征的综合模型具有卓越的诊断性能，在测试数据上的曲线下面积（AUC）为0.96（0.93 - 0.99）。该模型的准确率为0.91（0.85 - 0.95），灵敏度为0.88（0.73 - 0.97），特异度为0.92（0.85 - 0.96），阴性预测值为0.96（0.91 - 0.99），阳性预测值为0.77（0.61 - 0.89）。这个综合模型显著优于仅利用总体结节特征的模型（AUC = 0.85，P = 2.35e - 6），并且与仅基于内部特征的模型（AUC = 0.93，P = 1.01e - 1）具有相当的有效性。
结果支持了一种基于机器学习的方法的临床实用性，该方法整合了综合超声指标以可靠识别恶性PCTN。
