Noninvasive Prediction of Extrathyroidal Extension in Papillary Thyroid Carcinoma using Multiparametric MRI Radiomics and ADC Normalization

Background Gross extrathyroidal extension (ETE) drives surgical strategy and influences prognosis of papillary thyroid carcinoma (PTC). Precise, noninvasive assessment of gross ETE can tailor treatment to each patient. Purpose To evaluate the diagnosis performance of multiparametric magnetic resonance imaging (mpMRI)-based radiomics signature, combined with conventional quantitative MRI parameters, in predicting gross ETE in patients with PTC. Materials and Methods We retrospectively analyzed 140 PTC lesions imaged by mpMRI before surgery (March 2019 to November 2023). Radiomic features were extracted from T2-weighted imaging (T2WI), diffusion-weighted imaging (DWI), delayed contrast-enhanced (delayCE) images, and apparent diffusion coefficient (ADC) maps. Using six machine-learning algorithms, we built four single-modality models and two merged models. Univariate and multivariate logistic regression analyses were conducted on mpMRI quantitative parameters and optimal radiomic features, leading to the development of a nomogram model that incorporated independent predictive factors. The performance of the models was evaluated using receiver operating characteristic (ROC) curves, calibration curves, and decision curve analysis. The nomogram's efficacy was further assessed through 10-fold cross-validation, and the area under the curve (AUC) values were compared using the Delong test. Results Among the 140 PTC lesions, 37 (26.4%) exhibited gross ETE. The dataset was divided into a training cohort (102 cases, 1.5 Tesla MRI) and a test cohort (38 cases, 3.0 Tesla MRI). In the test cohort, the ExtraTrees-based merged model 2 (integrating T2WI, DWI, ADC, and delayCE) achieved the highest AUC of 0.853. The rad_signature (P = 0.005) and ADC_Best_rate (P < 0.001) emerged as independent predictors. The nomogram yielded AUCs of 0.893 (the training cohort) and 0.866 (the test cohort), with average 10-fold cross-validation AUCs of 0.908 and 0.853, respectively. The Hosmer-Lemeshow test confirmed the good fit (P = 0.704 and 0.533), and decision curve analysis suggested that the nomogram provided clinical benefit across a 0%-100% probability range in the training cohort and a 0%-50% range in the test cohort. Conclusion Our mpMRI-based nomogram model, integrating radiomics signature and ADC_Best_rate, effectively predicts gross ETE in PTC and offers a robust, noninvasive tool to guide surgical planning.