Development and validation of an ultrasound-based deep learning radiomics nomogram for risk assessment of lymph node metastasis in papillary thyroid microcarcinoma.

Background: Accurate preoperative prediction of cervical lymph node metastasis (LNM) in papillary thyroid microcarcinoma (PTMC) remains challenging, particularly because occult nodal disease is common and conventional ultrasound (US) assessment is operator-dependent. This study aimed to develop and validate a multicenter US-based deep learning radiomics nomogram integrating intratumoral, peritumoral, and clinical features for individualized LNM risk assessment in patients with PTMC. Methods: We retrospectively and prospectively collected multi-center data from three hospitals. Patients who underwent total thyroidectomy or lobectomy with lymph node dissection were allocated to a training set (763 cases), an external test set (118 cases), and a prospective validation set (94 cases). Radiomics features from within the tumor and deep learning (DL) features from the peritumoral region were extracted from the largest cross-sectional US image. Twelve machine learning (ML) models were built using the integrated features and evaluated on the test set to identify the optimal one. The machine learning prediction score (ML-score) was incorporated with clinical factors into a nomogram, and its performance was evaluated using area under the receiver operating characteristic curve (AUC), decision curve analysis (DCA), and clinical impact curves (CICs) analysis. Results: The support vector machine (SVM) model showed the best overall performance among the twelve ML algorithms and was used to generate the SVM-score. Multivariable logistic regression identified age, sex, maximum tumor diameter, genetic mutation status, and SVM-score as independent predictors of LNM. The integrated nomogram achieved AUCs of 0.894 in the training cohort, 0.842 in the external test cohort, and 0.856 in the prospective validation cohort. Calibration curves showed good agreement between predicted and observed LNM risk, with Hosmer-Lemeshow test P values of 0.189, 0.383, and 0.254 in the three cohorts, respectively. DCA and CIC demonstrated that the nomogram provided greater clinical net benefit than the clinical model or SVM-score model alone. Conclusions: The US-based DL radiomics nomogram integrating imaging features and clinical factors showed good performance for preoperative prediction of LNM in patients with PTMC. This non-invasive tool may assist individualized cervical lymph node risk stratification and support clinical decision-making.