超声组学标签对桥本甲状腺炎患者甲状腺结节恶性预测的意义
该单中心回顾性研究纳入110例桥本甲状腺炎患者的193个甲状腺结节(恶性110个、良性83个),从B型超声和超声造影的瘤内与瘤周区域提取超声组学标签构建模型。堆叠回归模型在训练集和测试集的AUC分别为0.997和0.867,敏感度96.6%和91.3%,特异度98.5%和81.2%,提示双区域双模态联合可提升诊断效能。
要点
- 堆叠回归模型测试集AUC为0.867,敏感度91.3%、特异度81.2%。
- 瘤周CEUS模型测试集AUC为0.742,瘤内CEUS模型为0.622,两者AUC差异显著(P<0.05)。
- 瘤内与瘤周B-US模型AUC差异无统计学意义(P>0.05)。
核心结论
- 超声组学方法可有效预测桥本甲状腺炎患者甲状腺结节的良恶性。
- 联合甲状腺双区域和双模态可提升超声组学模型的诊断效能。
研究设计与样本
- 单中心回顾性研究,纳入110例患者的193个甲状腺结节(恶性110个、良性83个)。
- 采集B型超声和超声造影图像,从瘤内和瘤周区域提取超声组学标签。
- 构建瘤内B-US、瘤周B-US、瘤内CEUS、瘤周CEUS超声组学模型及堆叠回归模型,以ROC曲线下面积比较诊断效能。
诊断标准与关键数值
- 训练集:瘤内B-US、瘤周B-US、瘤内CEUS、瘤周CEUS及堆叠回归模型的AUC(95%CI)分别为0.872(0.812,0.932)、0.815(0.747,0.882)、0.739(0.659,0.819)、0.890(0.836,0.943)、0.997(0.992,1.000)。
- 测试集:上述模型的AUC(95%CI)分别为0.799(0.650,0.948)、0.851(0.727,0.974)、0.622(0.440,0.805)、0.742(0.573,0.911)、0.867(0.741,0.992)。
- 训练集:上述模型的敏感度分别为82.8%、89.7%、71.3%、74.7%、96.6%,特异度分别为80.6%、58.2%、67.2%、91.0%、98.5%。
- 测试集:上述模型的敏感度分别为69.6%、78.3%、43.5%、78.3%、91.3%,特异度分别为93.8%、87.5%、93.3%、75.0%、81.2%。
- 堆叠回归模型与瘤内B-US、瘤周B-US模型相比AUC差异无统计学意义(P>0.05),瘤内CEUS与瘤周CEUS模型AUC差异显著(P<0.05)。
临床应用与阅片要点
- 基于超声组学标签的堆叠回归模型在测试集显示良好的校准和区分能力。
- 联合瘤内与瘤周区域、B-US与CEUS双模态的超声组学模型诊断效能优于单一区域或单一模态模型。
局限性
- 原文未报告局限性。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
英文原文(摘要)· 点击展开
RATIONALE AND OBJECTIVES: It remains a challenge to determine the nature of thyroid nodules (TNs) with Hashimoto's thyroiditis (HT). We aim to investigate the multiregional ultrasomics signatures obtained from B-mode ultrasound (B-US) and contrast-enhanced ultrasound (CEUS) images for predicting malignancy in TNs of patients with HT.
MATERIALS AND METHODS: B-US and CEUS images of 193 nodules (110 malignant and 83 benign nodules) from 110 patients were retrospectively collected in the single-center study, extracting ultrasomics signatures from the intratumoral (In) and peritumoral (Peri) regions of the thyroid. In-B-US, Peri-B-US, In-CEUS, and Peri-CEUS ultrasomics models and a stacking regression model were constructed, and the diagnostic performance of the models was evaluated by comparing the area under the receiver operating characteristic curve (ROC).
RESULTS: The In-B-US, Peri-B-US, In-CEUS, Peri-CEUS, and stacking regression model in the training and testing datasets which attained AUC (95% CI) of 0.872(0.812, 0.932), 0.815(0.747, 0.882), 0.739(0.659, 0.819), 0.890(0.836, 0.943), 0.997(0.992, 1.000) and 0.799(0.650, 0.948), 0.851(0.727, 0.974), 0.622(0.440, 0.805), 0.742(0.573, 0.911), 0.867(0.741, 0.992); sensitivity of 82.8%, 89.7%, 71.3%, 74.7%, 96.6% and 69.6%, 78.3%, 43.5%, 78.3%, 91.3%; specificity of 80.6%, 58.2%, 67.2%, 91.0%, 98.5% and 93.8%, 87.5%, 93.3%, 75.0%, 81.2%, respectively. The stacking regression model based on ultrasomics signatures showed favorable calibration and discriminative capabilities. Compared to the stacking regression model, the difference in AUC between the In-B-US and Peri-B-US models was not statistically significant (P > 0.05). However, the difference in AUC between the In-CEUS and Peri-CEUS models was significant (P < 0.05).
CONCLUSION: The application of an ultrasomics approach can effectively predict the benign or malignant nature of TNs accompanied by HT. The diagnostic performance of the ultrasomics model was improved by combining the dual-region and dual-mode of thyroid.
引用
- Sun M, Qu H, Xia H, et al. Implications of a Ultrasomics Signature for Predicting Malignancy in Thyroid Nodules with Hashimoto's Thyroiditis. Acad Radiol. 2024. PMID: 38796400. doi:10.1016/j.acra.2024.05.016
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。