Kim Jinyoung, Kim Min-Hee, Lim Dong-Jun, Lee Hankyeol, Lee Jae Jun, Kwon Hyuk-Sang, Kim Mee Kyoung, Song Ki-Ho, Kim Tae-Jung, Jung So Lyung, Lee Yong Oh, Baek Ki-Hyun
Division of Endocrinology and Metabolism, Department of Internal Medicine, College of Medicine, The Catholic University of Korea, Seoul, Korea.
Department of Computer Engineering, Hongik University, Seoul, Korea.
Endocrinol Metab (Seoul). 2025 Apr;40(2):216-224. doi: 10.3803/EnM.2024.2058. Epub 2025 Jan 13.
This study aimed to evaluate the applicability of deep learning technology to thyroid ultrasound images for classification of thyroid nodules.
This retrospective analysis included ultrasound images of patients with thyroid nodules investigated by fine-needle aspiration at the thyroid clinic of a single center from April 2010 to September 2012. Thyroid nodules with cytopathologic results of Bethesda category V (suspicious for malignancy) or VI (malignant) were defined as thyroid cancer. Multiple deep learning algorithms based on convolutional neural networks (CNNs) -ResNet, DenseNet, and EfficientNet-were utilized, and Siamese neural networks facilitated multi-view analysis of paired transverse and longitudinal ultrasound images.
Among 1,048 analyzed thyroid nodules from 943 patients, 306 (29%) were identified as thyroid cancer. In a subgroup analysis of transverse and longitudinal images, longitudinal images showed superior prediction ability. Multi-view modeling, based on paired transverse and longitudinal images, significantly improved the model performance; with an accuracy of 0.82 (95% confidence intervals [CI], 0.80 to 0.86) with ResNet50, 0.83 (95% CI, 0.83 to 0.88) with DenseNet201, and 0.81 (95% CI, 0.79 to 0.84) with EfficientNetv2_ s. Training with high-resolution images obtained using the latest equipment tended to improve model performance in association with increased sensitivity.
CNN algorithms applied to ultrasound images demonstrated substantial accuracy in thyroid nodule classification, indicating their potential as valuable tools for diagnosing thyroid cancer. However, in real-world clinical settings, it is important to aware that model performance may vary depending on the quality of images acquired by different physicians and imaging devices.
本研究旨在评估深度学习技术在甲状腺超声图像中对甲状腺结节进行分类的适用性。
这项回顾性分析纳入了2010年4月至2012年9月在单中心甲状腺诊所接受细针穿刺检查的甲状腺结节患者的超声图像。细胞病理学结果为贝塞斯达分类V类（可疑恶性）或VI类（恶性）的甲状腺结节被定义为甲状腺癌。使用了基于卷积神经网络（CNN）的多种深度学习算法——ResNet、DenseNet和EfficientNet，连体神经网络促进了对成对横向和纵向超声图像的多视图分析。
在来自943例患者的1048个分析的甲状腺结节中，306个（29%）被确定为甲状腺癌。在横向和纵向图像的亚组分析中，纵向图像显示出更好的预测能力。基于成对横向和纵向图像的多视图建模显著提高了模型性能；ResNet50的准确率为0.82（95%置信区间[CI]，0.80至0.86），DenseNet201的准确率为0.83（95%CI，0.83至0.88），EfficientNetv2_s的准确率为0.81（95%CI，0.79至0.84）。使用最新设备获得的高分辨率图像进行训练，随着敏感性增加，往往会提高模型性能。
应用于超声图像的CNN算法在甲状腺结节分类中显示出较高的准确性，表明它们作为诊断甲状腺癌的有价值工具的潜力。然而，在实际临床环境中，重要的是要意识到模型性能可能因不同医生和成像设备获取的图像质量而异。
