Lu W J, Qiu Y R, Wu Y W, Li J, Chen R, Chen S N, Lin Y Y, OuYang L Y, Chen J Y, Chen F, Qiu S D
The Second Affiliated Hospital of Guangzhou Medical University - Ultrasound.
The Second Clinical School of Guangzhou Medical University - Department of Clinical Medicine, Guangzhou, Guangdong, China.
Acta Endocrinol (Buchar). 2022 Oct-Dec;18(4):407-416. doi: 10.4183/aeb.2022.407.
To evaluate the diagnostic performance of radiomics features of two-dimensional (2D) and three-dimensional (3D) ultrasound (US) in predicting extrathyroidal extension (ETE) status in papillary thyroid carcinoma (PTC).
2D and 3D thyroid ultrasound images of 72 PTC patients confirmed by pathology were retrospectively analyzed. The patients were assigned to ETE and non-ETE. The regions of interest (ROIs) were obtained manually. From these images, a larger number of radiomic features were automatically extracted. Lastly, the diagnostic abilities of the radiomics models and a radiologist were evaluated using receiver operating characteristic (ROC) analysis. We extracted 1693 texture features firstly.
The area under the ROC curve (AUC) of the radiologist was 0.65. For 2D US, the mean AUC of the three classifiers separately were: 0.744 for logistic regression (LR), 0.694 for multilayer perceptron (MLP), 0.733 for support vector machines (SVM). For 3D US they were 0.876 for LR, 0.825 for MLP, 0.867 for SVM. The diagnostic efficiency of the radiomics was better than radiologist. The LR model had favorable discriminate performance with higher area under the curve.
Radiomics based on US image had the potential to preoperatively predict ETE. Radiomics based on 3D US images presented more advantages over radiomics based on 2D US images and radiologist.
评估二维（2D）和三维（3D）超声（US）的影像组学特征在预测甲状腺乳头状癌（PTC）甲状腺外侵犯（ETE）状态方面的诊断性能。
回顾性分析72例经病理证实的PTC患者的2D和3D甲状腺超声图像。将患者分为ETE组和非ETE组。手动获取感兴趣区域（ROI）。从这些图像中自动提取大量影像组学特征。最后，使用受试者操作特征（ROC）分析评估影像组学模型和放射科医生的诊断能力。我们首先提取了1693个纹理特征。
放射科医生的ROC曲线下面积（AUC）为0.65。对于2D US，三个分类器的平均AUC分别为：逻辑回归（LR）为0.744，多层感知器（MLP）为0.694，支持向量机（SVM）为0.733。对于3D US，它们分别为：LR为0.876，MLP为0.825，SVM为0.867。影像组学的诊断效率优于放射科医生。LR模型具有良好的鉴别性能，曲线下面积更高。
基于超声图像的影像组学有术前预测ETE的潜力。基于3D US图像的影像组学比基于2D US图像的影像组学和放射科医生具有更多优势。
