Radiomics based on arterial-venous mixed images derived from dual-energy CT data in diagnosis of lymph nodes metastasis of papillary thyroid cancer
10.3760/cma.j.cn112149-20210421-00395
- VernacularTitle:双能CT动静脉期加权融合图像影像组学对甲状腺乳头状癌颈部淋巴结转移的诊断价值
- Author:
Yan ZHOU
1
;
Xiaoquan XU
;
Guoyi SU
;
Xinwei TAO
;
Yingqian GE
;
Yan SI
;
Meiping SHEN
;
Feiyun WU
Author Information
1. 南京医科大学第一附属医院放射科 210029
- Keywords:
Thyroid neoplasms;
Carcinoma, papillary;
Lymphatic metastasis;
Radiomics;
Tomography, X-ray computed
- From:
Chinese Journal of Radiology
2021;55(7):703-709
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To explore the diagnostic value of radiomics based on arterial-venous mixed images derived from dual-energy CT (DECT) data in diagnosis of cervical lymph nodes (LNs) metastasis of papillary thyroid cancer (PTC).Methods:From June 2017 to December 2018, eighty-four patients with preoperatively DECT scanning and pathologically confirmed PTC (129 non-metastatic LNs and 97 metastatic LNs) in the First Affiliated Hospital of Nanjing Medical University were included in this study. The clinical and imaging data of all patients were retrospectively analyzed. The training cohort consisted of 62 PTC cases with 156 LNs (91 non-metastatic LNs and 65 metastatic LNs). An independent validation cohort consisted of 22 PTC patients with 70 LNs (38 non-metastatic LNs and 32 metastatic LNs). Semi-automatic LNs segmentation was conducted on arterial-venous mixed images derived from DECT using Syngo.via Frontier Radiomics software. Totally 1 226 radiomics features were extracted from arterial-venous mixed images for each LN. The least absolute shrinkage and selection operator (LASSO) regression was applied for radiomics features selection and signature building. The logistic regression modeling was used to construct diagnostic models based on the CT image features of LNs (model 1), the radiomics signature (model 2) and the combination of the CT image features and radiomics signature (model 3). An intuitive nomogram was plotted for model 3. The ROC curve analyses and area under the curve (AUC) were performed to evaluate the diagnostic efficiency of the three models, with the performances compared using the Delong test.Results:Model 1 was developed with LNs shape, degree of enhancement, pattern of enhancement, calcification and extra nodal extension. Three arterial phase radiomics features were selected and used to establish radiomics signature using LASSO regression (model 2). Model 3 was developed with LNs size, shape, degree of enhancement and radiomics signature. In both the training and validation cohort, model 3 showed the best diagnostic performance (AUC=0.965, 0.933), followed by model 2 (AUC=0.947, 0.910), and both these two models significantly outperformed model 1 (AUC=0.850, 0.846) (training cohort, Z=4.066 and 3.758, P both<0.001; validation cohort, Z=2.871 and 1.998, P=0.017 and 0.042) respectively. Conclusion:The radiomics model based on arterial-venous mixed images derived from DECT data can realize effective diagnosis of LNs metastasis in patients with PTC; and the combination model of radiomics signature with CT image features can further improve the diagnostic accuracy.