LAUSR

Objective: To explore extrathyroidal extension (ETE) in children and adolescents with papillary thyroid carcinoma using a multiclassifier ultrasound radiomic model. Methods: In this study, data from 164 pediatric patients with papillary thyroid cancer (PTC) were retrospectively analyzed and patients were randomly divided into a training cohort (115) and a validation cohort (49) in a 7:3 ratio. To extract radiomics features from ultrasound images of the thyroid, areas of interest (ROIs) were delineated layer by layer along the edge of the tumor contour. The feature dimension was then reduced using the correlation coefficient screening method, and 16 features with a nonzero coefficient were chosen using Lasso. Then, in the training cohort, four supervised machine learning radiomics models (k-nearest neighbor, random forest, support vector machine [SVM], and LightGBM) were developed. ROC and decision-making curves were utilized to compare model performance, which was validated using validation cohorts. In addition, the SHapley Additive exPlanations (SHAP) framework was applied to explain the optimal model. Results: In the training cohort, the average area under the curve (AUC) was 0.880 (0.835-0.927), 0.873 (0.829-0.916), 0.999 (0.999-1.000), and 0.926 (0.892-0.926) for the SVM, KNN, random forest, and LightGBM, respectively. In the validation cohort, the AUC for the SVM was 0.784 (0.680-0.889), for the KNN, it was 0.720 (0.615-0.825), for the random forest, it was 0.728 (0.622-0.834), and for the LightGBM, it was 0.832 (0.742-0.921). Generally, the LightGBM model performed well in both the training and validation cohorts. From the SHAP results, original_shape_MinorAxisLength,original_shape_Maximum2DDiameterColumn, and wavelet-HHH_glszm_SmallAreaLowGrayLevelEmphasis have the most significant effect on the model. Conclusions: Our combined model based on machine learning and ultrasonic radiomics demonstrate the excellent predictive ability for extrathyroidal extension (ETE) in pediatric PTC.