Feasibility study of predicting thyroid papillary carcinoma central lymph node metastasis based on wavelet texture analysis using venous phase CT images
10.3760/cma.j.issn.1005-1201.2019.11.004
- VernacularTitle: 基于甲状腺乳头状癌结节CT静脉期图像小波纹理分析预测中央组淋巴结转移的可行性研究
- Author:
Shasha SHEN
1
;
Dan HAN
1
;
Lun ZHAO
2
;
Xiuli LI
2
;
Zhenghua ZHANG
1
;
Wen ZHAO
1
;
Wenyan WEI
1
Author Information
1. Department of Radiology, the First Affiliated Hospital of Kunming Medical University, Kunming 650032, China
2. Deepwise Artificial Intelligence Lab, Beijing 100080, China
- Publication Type:Journal Article
- Keywords:
Thyroid neoplasms;
Neoplasm metastasis;
Tomography, spiral computed;
Texture analysis
- From:
Chinese Journal of Radiology
2019;53(11):946-951
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To investigate the value of CT wavelet texture analysis based on primary tumor of papillary thyroid carcinoma (PTC) in predicting central lymph node metastasis (CLNM).
Methods:A retrospective analysis was performed to 250 patients (307 nodules) who pathologically confirm with PTC in the First Affiliated Hospital of Kunming Medical University from December 2013 to August 2019. Thyroid dual phase scanning was performed in all patients within two weeks before surgery. All patients underwent central or total cervical lymph node dissection, among which 160 cases (189 nodules) were classified as training sets, while 90 cases (118 nodules) were in the verification sets. Besides, all patients were divided into CLNM group and no CLNM group according to pathology. The DeepWise software were used to manually delineate PTC primary nodules on venous phase CT images, and 576 wavelet texture features were extracted. The differences of texture feature parameters between the two groups were compared. The top 10 wavelet texture features of the area under curve (AUC) value were manually selected as the best parameters. Multivariable logistic regression analysis was used to establish and verify the model, the optimal cutoff value was found by using receiver operating characteristic curve analysis.
Results:Totally 124 features were statistical difference between the two groups. The top 10 characteristic parameters for manual diagnosis with AUC values ranged from 0.599 to 0.630 (P<0.05), multi-collinearity test and multi-logstic regression analysis showed that there was no collinear correlation between the above 10 features, and small-area low-gray emphasis was an independent predictor of risk factors. The AUC value, sensitivity, specificity, and accuracy of the predictive model for the diagnosis of CLNM in the training set were 0.693, 62.84%, 60.47%, 62.96% and validation set were 0.602, 64.95%, 33.33%, and 59.32%, respectively.
Conclusion:Wavelet texture analysis in CT venous phase may allow detection of CLNM of PTC.
