Cozzi Joseph L, Li Hui, Fuhrman Jordan D, Lan Li, Williams Jelani, Finnerty Brendan, Fahey Thomas J, Tumati Abhinay, Genender Joshua, Keutgen Xavier M, Giger Maryellen L
Department of Radiology, University of Chicago, Chicago, IL, USA.
Division of General Surgery and Surgical Oncology, Department of Surgery, University of Chicago Medicine, Endocrine Surgery Research Program, Chicago, IL, USA.
Int J Comput Assist Radiol Surg. 2025 Feb;20(2):259-267. doi: 10.1007/s11548-024-03294-w. Epub 2025 Jan 3.
Thyroid nodules are common, and ultrasound-based risk stratification using ACR's TIRADS classification is a key step in predicting nodule pathology. Determining thyroid nodule contours is necessary for the calculation of TIRADS scores and can also be used in the development of machine learning nodule diagnosis systems. This paper presents the development, validation, and multi-institutional independent testing of a machine learning system for the automatic segmentation of thyroid nodules on ultrasound.
The datasets, containing a total of 1595 thyroid ultrasound images from 520 patients with thyroid nodules, were retrospectively collected under IRB approval from University of Chicago Medicine (UCM) and Weill Cornell Medical Center (WCMC). Nodules were manually contoured by a team of UCM and WCMC physicians for ground truth. An AttU-Net, a U-Net architecture with additional attention weighting functions, was trained for the segmentations. The algorithm was validated through fivefold cross-validation by nodule and was tested on two independent test sets: one from UCM and one from WCMC. Dice similarity coefficient (DSC) and percent Hausdorff distance (%HD), Hausdorff distance reported as a percent of the nodule's effective diameter, served as the performance metrics.
On multi-institutional independent testing, the AttU-Net yielded average DSCs (std. deviation) of 0.915 (0.04) and 0.922 (0.03) and %HDs (std. deviation) of 12.9% (4.6) and 13.4% (6.3) on the UCM and WCMC test sets, respectively. Similarity testing showed the algorithm's performance on the two institutional test sets was equivalent up to margins of DSC 0.013 and %HD 1.73%.
This work presents a robust automatic thyroid nodule segmentation algorithm that could be implemented for risk stratification systems. Future work is merited to incorporate this segmentation method within an automatic thyroid classification system.
甲状腺结节很常见，使用美国放射学会（ACR）的甲状腺影像报告和数据系统（TIRADS）分类进行基于超声的风险分层是预测结节病理的关键步骤。确定甲状腺结节轮廓对于计算TIRADS评分是必要的，并且还可用于机器学习结节诊断系统的开发。本文介绍了一种用于超声甲状腺结节自动分割的机器学习系统的开发、验证和多机构独立测试。
在芝加哥大学医学中心（UCM）和威尔康奈尔医学院（WCMC）的机构审查委员会（IRB）批准下，回顾性收集了包含来自520例甲状腺结节患者的总共1595张甲状腺超声图像的数据集。结节由UCM和WCMC的一组医生手动勾勒轮廓作为真实标准。使用带有附加注意力加权函数的U-Net架构AttU-Net进行分割训练。该算法通过结节的五重交叉验证进行验证，并在两个独立测试集上进行测试：一个来自UCM，一个来自WCMC。骰子相似系数（DSC）和百分比豪斯多夫距离（%HD，豪斯多夫距离以结节有效直径的百分比报告）用作性能指标。
在多机构独立测试中，AttU-Net在UCM和WCMC测试集上的平均DSC（标准差）分别为0.915（0.04）和0.922（0.03），%HD（标准差）分别为12.9%（4.6）和13.4%（6.3）。相似性测试表明，该算法在两个机构测试集上的性能在DSC相差0.013和%HD相差1.73%的范围内相当。
这项工作提出了一种强大的甲状腺结节自动分割算法，可用于风险分层系统。未来有必要将这种分割方法纳入自动甲状腺分类系统。
