Development and validation of a machine learning model integrating spectral computed tomography-derived three‑dimensional quantitative parameters and clinical features for predicting minimal extrathyroidal extension in papillary thyroid microcarcinoma.

Background: Preoperative identification of minimal extrathyroidal extension (mETE) in papillary thyroid microcarcinoma (PTMC) remains clinically challenging due to the limited sensitivity and specificity of conventional imaging. Although spectral computed tomography (CT) enables three‑dimensional (3D) quantitative tissue characterization, its integration with clinical data for mETE prediction has yet to be systematically evaluated. This study aimed to develop and validate multiple machine learning (ML) models integrating spectral CT 3D parameters with clinical features to predict mETE in PTMC patients. Methods: A total of 253 patients with pathologically proven PTMC who had undergone spectral CT were retrospectively enrolled in this study and randomly assigned to either a training cohort (n=177) or a testing cohort (n=76). Patients with prior thyroid surgery, poor image quality, missing clinical data, or other malignancies were excluded. We collected 12 clinical features along with 13 spectral CT 3D quantitative parameters. mETE status was confirmed by postoperative histopathology. Feature selection was performed using Pearson correlation analysis and the Boruta algorithm. Predictive models were developed using 4 ML algorithms: light gradient boosting machine (LightGBM), multilayer perceptron classifier (MLP-C), Adaptive Boosting (AdaBoost) and logistic regression (LR). We evaluated the predictive models using receiver operating characteristic (ROC) curve analysis. The accuracy, area under the curve (AUC), sensitivity, specificity, positive predictive value and negative predictive value were all calculated. Shapley additive explanations (SHAP) values were used to interpret variable contributions. Results: A total of 253 patients (170 mETE-positive, 83 mETE-negative; mean age 39.77±10.71 years) were included. Four variables were selected for inclusion in the final model, namely 3D-arterial enhancement fraction (AEF), BRAF V600E status, capsular contact status, 3D-arterial phase (AP)40keV. LightGBM, MLP-C, AdaBoost, and LR model demonstrated satisfactory diagnostic efficacy, with the AUCs of 0.99, 0.95, 0.91, 0.88 and 0.78, 0.77, 0.81, 0.80 in the training and testing cohorts, respectively. The SHAP method showed that the top three variables ranked by contribution degree were: 3D-AEF, BRAF V600E status, capsular contact status. Conclusions: ML models that incorporate spectral CT 3D parameters and clinical features offer a promising approach for preoperative prediction of mETE in PTMC. Following external validation in multicenter cohorts, these models may serve as adjunctive tools to inform clinical decision-making and risk stratification.