Radiomics Analysis of [18F]-Fluorodeoxyglucose-Avid Thyroid Incidentalomas Improves Risk Stratification and Selection for Clinical Assessment.

Abstract

Background: [18F]-Fluorodeoxyglucose (FDG)-avid thyroid lesions incidentally detected on positron emission tomography/computed tomography (PET/CT) scans represent a tumor lesion in about 30% of cases. The present study evaluated the ability of PET metrics and radiomics features to predict final diagnosis of [18F]FDG thyroid incidentalomas (TIs). Methods: A total of 104 patients with 107 TIs were retrospectively studied; 30 nodules (28%) were diagnosed as malignant. After volumetric segmentation of each thyroid lesion, metabolic tumor volume (MTV), total lesion glycolysis (TLG), standardized uptake values (SUVs), and metabolic heterogeneity were estimated, and 107 radiomics features were extracted following a standard protocol. Results: MTV, TLG, SUVmax, SUVmean, and SUVpeak among functional PET parameters, and gray-level co-occurrence matrix (GLCM)_InverseDifferenceMoment, shape_Sphericity, GLCM_SumSquares, firstorder_Maximum2DDiameterSlice, firstorder_Energy, and GLCM_Contrast among nonredundant radiomics features, showed significantly different values between malignant and benign TIs (Mann-Whitney U-test, p < 0.01 for all). Univariate logistic regression revealed that these parameters demonstrated good ability to predict final diagnosis of TIs (p < 0.02 for all). Shape_Sphericity was the best predictor classifying 82% of TIs correctly (p < 0.0001). Only TLG, SUVmax, and shape_Sphericity retained significance (p < 0.0001) by multivariate analysis. Malignant lesion prevalence increased from 7% to 100% in accordance with the number (score, 0-3) of the three positive parameters present (χ2 trend, p < 0.0001). A score of 0 excludes malignant TIs with a negative predictive value of 93%, while a score of 3 predicted malignancy with a positive predictive value of 100%. Conclusions: PET metrics and radiomics analysis can improve identification of [18F]FDG-avid TIs at high risk of malignancy. A model based on TLG, SUVmax, and shape_Sphericity may allow prediction of a final diagnosis, providing useful information for the management of TIs.