Abstract
BACKGROUND: In recent years, the study of radiomics in thyroid diseases has developed rapidly. This study aimed to establish a preoperative radiomics prediction model for central compartment lymph node metastases (CLNMs) in papillary thyroid microcarcinoma (PTMC) patients using gradient-boosting decision tree (GBDT) model and evaluate the performance of the model. METHODS: A total of 274 patients with PTMC admitted for thyroid surgery at China-Japan Union Hospital of Jilin University from January 2020 to July 2022 were retrospectively analyzed. Patients were randomized into training and validation cohorts according to a ratio of 8:2. Radiomics features were extracted from the ultrasound (US) images of PTMC lesions. The open-source software Pyradiomics was used to extract radiomics features, and WEKA software was used to select CLNM-related radiomics features. Clinical risk factors for CLNM were screened by statistical methods. The GBDT model was constructed by combining radiomics features and clinical risk factors, and compared with the diagnostic efficacy of US-reported cervical lymph node status. Shapley Additive exPlanations (SHAP) was applied to visualize and analyze the GBDT model globally and locally. RESULTS: A total of seven radiomics features were significantly correlated with central lymph node status in the training and validation cohorts. The predictors in the GBDT model included the radiomics features, sex, age, and body mass index (BMI). The area under the curve (AUC) values of the GBDT model in the training and validation cohorts were 0.946 [95% confidence interval (CI): 0.920-0.972] and 0.845 (95% CI: 0.714-0.976), respectively, compared with 0.583 (95% CI: 0.508-0.659) and 0.582 (95% CI: 0.430-0.736) for US-reported lymph node status alone. The Delong test showed a significant difference between AUS in the training and validation cohorts (P<0.001, respectively). SHAP visual analysis showed the effect of each parameter on the GBDT model globally and locally. Decision curve analysis demonstrated the clinical utility of the GBDT model. CONCLUSIONS: The prediction of CLNM by the GBDT model, based on US radiomics features and clinical factors, can be better than that by using US alone in patients with PTMC. Furthermore, the GBDT model may serve a guidance of clinical decision for patient's treatment strategy.