Radiomics and deep learning for large volume lymph node metastasis in papillary thyroid carcinoma.

Background: Thyroid cancer is prone to early lymph node metastasis (LNM), and patients with large volume LNM (LVLNM) tend to have a poorer prognosis. The aim of this study was to predict LVLNM in before surgery based on radiomics and deep learning (DL).

Methods: A multicenter retrospective study was performed, including 854 papillary thyroid carcinoma (PTC) patients from three centers. Radiomics features were extracted. Logistic regression (LR), support vector machine (SVM), K-nearest neighbors (KNN), multi-layer perceptron (MLP), random forest (RF), ExtraTrees, extreme gradient boosting (XGBoost), and light gradient boosting machine (LightGBM) algorithms were used to construct radiomics models. AlexNet, DenseNet121, inception_v3, ResNet50, and transformer algorithms were used to construct DL models. The receiver operating characteristic (ROC) curve was employed to select the better-performing model. A combined model was then created by merging radiomics features and DL features. The least absolute shrinkage and selection operator (LASSO) method was utilized to identify metabolites and radiomics features with non-zero coefficients. The performance of the models was evaluated using area under the curve (AUC), accuracy (ACC), sensitivity (SEN), specificity (SPE), positive predictive value (PPV), negative predictive value (NPV), and F1-score.

Results: A total of 1,357 radiomics features were extracted. Among the radiomics models, the ExtraTrees model demonstrated the optimal diagnostic capabilities with an AUC of 0.787 [95% confidence interval (CI): 0.715-0.858], and DenseNet121 DL model demonstrated the optimal diagnostic capabilities with an AUC of 0.766 (95% CI: 0.683-0.848). Furthermore, the combined model, named the Thy-DL-Radiomics model, exhibited an AUC of 0.839 (95% CI: 0.758-0.920) in the internal validation set and 0.789 (95% CI: 0.718-0.859) in the external validation set.

Conclusions: A radiomics-DL features integrated model can predict LVLNM in PTC patients and provide guidance for personalized treatment.