Machine learning-based multiparametric MRI radiomics for predicting the aggressiveness of papillary thyroid carcinoma

Abstract

PurposeTo investigate the predictive capability of machine learning-based multiparametric magnetic resonance (MR) imaging radiomics for evaluating the aggressiveness of papillary thyroid carcinoma (PTC) preoperatively. MethodsThis prospective study enrolled consecutive patients who underwent neck MR scans and subsequent thyroidectomy during the study interval. The diagnosis and aggressiveness of PTC were determined by pathological evaluation of thyroidectomy specimens. Thyroid nodules were segmented manually on the MR images, and radiomic features were then extracted. Predictive machine learning modelling was used to evaluate the prediction of PTC aggressiveness. Area under the receiver operating characteristic curve (AUC) values for the model performance were obtained for radiomic features, clinical characteristics, and combinations of radiomic features and clinical characteristics. ResultsThe study cohort included 120 patients with pathology-confirmed PTC (training cohort: n = 96; testing cohort: n = 24). A total of 1393 features were extracted from T2-weighted, apparent diffusion coefficient (ADC) and contrast-enhanced T1-weighted MR images for each patient. The combination of Least Absolute Shrinkage and Selection Operator for radiomic feature selection and Gradient Boosting Classifier for classifying PTC aggressiveness achieving the AUC of 0.92. In contrast, clinical characteristics alone poorly predicted PTC aggressiveness, with an AUC of 0.56. ConclusionsOur study showed that machine learning-based multiparametric MR imaging radiomics could accurately distinguish aggressive from non-aggressive PTC preoperatively. This approach may be helpful for informing treatment strategies and prognosis of patients with aggressive PTC.