Department of Thyroid Surgery.
Big Data and Artificial Intelligence Laboratory.
Int J Surg. 2023 Nov 1;109(11):3337-3345. doi: 10.1097/JS9.0000000000000660.
Preoperative evaluation of the metastasis status of lateral lymph nodes (LNs) in papillary thyroid cancer is challenging. Strategies for using deep learning to diagnosis of lateral LN metastasis require additional development and testing. This study aimed to build a deep learning-based model to distinguish benign lateral LNs from metastatic lateral LNs in papillary thyroid cancer and test the model's diagnostic performance in a real-world clinical setting.
This was a prospective diagnostic study. An ensemble model integrating a three-dimensional residual network algorithm with clinical risk factors available before surgery was developed based on computed tomography images of lateral LNs in an internal dataset and validated in two external datasets. The diagnostic performance of the ensemble model was tested and compared with the results of fine-needle aspiration (FNA) (used as the standard reference method) and the diagnoses made by two senior radiologists in 113 suspicious lateral LNs in patients enrolled prospectively.
The area under the receiver operating characteristic curve of the ensemble model for diagnosing suspicious lateral LNs was 0.829 (95% CI: 0.732-0.927). The sensitivity and specificity of the ensemble model were 0.839 (95% CI: 0.762-0.916) and 0.769 (95% CI: 0.607-0.931), respectively. The diagnostic accuracy of the ensemble model was 82.3%. With FNA results as the criterion standard, the ensemble model had excellent diagnostic performance ( P =0.115), similar to that of the two senior radiologists ( P =1.000 and P =0.392, respectively).
A three-dimensional residual network-based ensemble model was successfully developed for the diagnostic assessment of suspicious lateral LNs and achieved diagnostic performance similar to that of FNA and senior radiologists. The model appears promising for clinical application.
术前评估甲状腺乳头状癌侧颈部淋巴结（LNs）的转移状态具有挑战性。使用深度学习策略诊断侧颈淋巴结转移需要进一步开发和测试。本研究旨在构建一种基于深度学习的模型，以区分甲状腺乳头状癌中良性侧 LNs 与转移性侧 LNs，并在真实临床环境中测试该模型的诊断性能。
这是一项前瞻性诊断研究。在内部数据集的侧颈淋巴结 CT 图像的基础上，开发了一种集成三维残差网络算法和术前临床危险因素的集成模型，并在两个外部数据集进行验证。该集成模型的诊断性能在 113 例可疑侧颈淋巴结中进行了测试，并与细针抽吸（FNA）（作为标准参考方法）和两名高级放射科医生的诊断结果进行了比较。
该集成模型诊断可疑侧颈淋巴结的受试者工作特征曲线下面积为 0.829（95%CI：0.732-0.927）。该集成模型的敏感性和特异性分别为 0.839（95%CI：0.762-0.916）和 0.769（95%CI：0.607-0.931），诊断准确率为 82.3%。以 FNA 结果为标准，该集成模型具有良好的诊断性能（P=0.115），与两名高级放射科医生相似（P=1.000 和 P=0.392）。
成功开发了一种基于三维残差网络的集成模型，用于可疑侧颈淋巴结的诊断评估，其诊断性能与 FNA 和高级放射科医生相当。该模型具有良好的临床应用前景。
