Prognostic Analysis of 131I Efficacy After Papillary Thyroid Carcinoma Surgery Based on CT Radiomics.

Abstract

To develop and validate a radiomics-clinical combined model combining preoperative computed tomography (CT) and clinical data from patients with papillary thyroid carcinoma (PTC) to predict the efficacy of initial postoperative 131I treatment. A total of 181 patients with PTC who received total thyroidectomy and initial 131I treatment were divided into training and testing sets (7:3 ratio). Univariate analysis and multivariate logistic regression were used to screen clinical factors affecting the therapeutic response to 131I treatment and construct a clinical model. Radiomics features extracted from preoperative CT images of PTCs were dimensionally reduced through recursive feature elimination and least absolute shrinkage and selection operator. Logistic regression was used to establish a radiomics model, and a radiomics-clinical combined model was developed by integrating the clinical model. The area under the curve (AUC), sensitivity, and specificity were used to evaluate the prediction performance of each model. Multivariate analysis revealed that pre-131I treatment serum thyroglobulin was an independent clinical risk factor affecting the efficacy of initial 131I treatment (P = .002), and the AUC, sensitivity, and specificity for predicting the efficacy of initial 131I treatment were 0.895, 0.899, and 0.816, respectively. After dimensionality reduction, 14 key CT radiomics features of PTCs were included. The established radiomics model predicted the efficacy of 131I treatment in the training and testing sets with AUCs of 0.825 and 0.809, sensitivities of 0.828 and 0.636, and specificities of 0.745 and 0.944, respectively. The combined model improved the AUC, sensitivity, and specificity in both sets. The preoperative CT-based radiomics model can effectively predict the efficacy of initial postoperative 131I treatment in patients with intermediate- or high-risk PTC, and the radiomics-clinical combined model exhibits better predictive performance.