Application of CT Texture Feature Analysis in Predicting Immunohistochemistry of Unidentifiable Thyroid Nodules

Abstract

The purpose of this study was to develop and validate a radiomics model for evaluating immunohistochemistry (IHC) in patients with unidentifiable thyroid nodules. 103 patients (training cohort: validation cohort about 3:1) with unidentifiable thyroid nodules who had undergone thyroidectomy and IHC were enrolled. The IHC parameters include cytokeratin-19 (CK-19), thyroperoxidase (TPO) and high molecular weight cytokeratin(34βE12). All the patients underwent CT scans before surgery and 3D slicer was used to feature extraction. Test-retest (concordance correlation coefficient, CCC) and spearman`s correlation coefficient (ρ)was used to select reproducible and nonredundant features. Kruskal-Wallis test(p<0.05) was used for feature selection, whereupon a feature-based model was built with support vector machine (SVM) methods. The performance of the radiomic models were assessed by accuracy (ACC), sensitivity (SEN), specificity (SPE), corresponding area under the curves (AUC), and independent validation. 86 reproducible and nonredundant features selected form 841 features were used for model building. The best performance of CK-19 model yielded an ACC of 84.4% (SEN: 0.93, SPE: 0.73, AUC: 0.87) in the training cohort and 80% (SEN: 0.75, SPE: 0.875) in validation cohort. The TPO predictive model yielded an ACC of 81.4% (SEN: 0.86, SPE:0.75, AUC: 0.84) in the training cohort, and 84.21% (SEN: 0.9, SPE: 0.778) in the validation cohort, respectively. However, the performance of 34βE12 predictive model was not significant (ACC: 65.7%), and had no validation. A radiomics model with excellent performance was developed for individualized, non-invasive prediction of CK-19 and TPO based on CT images which may be used to identify benign and malignant thyroid nodules.