Development and internal validation of a multivariable MRI-based radiomics prediction model for preoperative cervical lymph node metastasis in thyroid cancer.

BACKGROUND: Accurate preoperative assessment of cervical lymph node metastasis (LNM) is crucial for surgical planning in thyroid cancer (TC), yet conventional ultrasound and computed tomography have limited sensitivity. This study aimed to develop and internally validate a magnetic resonance imaging (MRI)-based radiomics prediction model for preoperative LNM status in patients with TC. METHODS: We prospectively enrolled 97 consecutive patients with pathologically confirmed TC who underwent preoperative MRI and cervical lymph node dissection at The First Affiliated Hospital of Guangxi Medical University from April 2017 to January 2019. The reference standard was histopathology. Patients were randomly divided into training and validation sets at a 7:3 ratio. A total of 1,691 radiomics features were extracted from T2-weighted imaging (T2WI) and T1-weighted heterogeneous contrast-enhanced imaging (T1C+), respectively. Using the Boruta algorithm, we selected a subset of optimized features for the radiomics signature, and a random forest algorithm was used to construct the radiomics model. Subsequently, the ability of this model to predict the status of lymph nodes of patients with TC was evaluated through receiver operating characteristic (ROC) curve analysis. RESULTS: The combined model had better diagnostic performance than did models based on T2WI and T1C+ alone. Of the 97 patients, 61 (62.9%) were LNM-positive. In the training set, the area under the curve (AUC) values for the T2WI, T1C+, and combined models were 0.914, 0.959, and 0.963, respectively; the corresponding AUC values in the validation set were 0.853, 0.881, and 0.900, respectively. In the validation set, the combined model achieved an AUC of 0.900, with a sensitivity of 0.944 and specificity of 0.500. CONCLUSIONS: The MRI based radiomics model shows promising discriminative ability for preoperative LNM prediction in TC, but its low specificity (0.500) limits standalone clinical utility. External validation in larger cohorts is required before clinical implementation.