Comparison of different machine learning models based on ultrasound-based radiomics to predict central lymph node metastasis of papillary thyroid carcinoma

Abstract

Abstract Background: Accurate methods to predict central lymph node metastases preoperatively are needed to improve the management of patients with papillary thyroid carcinoma. The objective of this study was to apply machine learning models based on ultrasound radiomic data to predict central lymph node metastases and to identify the best differential diagnosis model. Methods: Clinicopathological information was retrospectively collected. All patients underwent preoperative thyroid ultrasound and postoperative lymph node pathology analysis. The regions of interest were manually drawn using a three-dimensional slicer and features specific to each area of injury were extracted. Five machine learning models were established to identify the appearance of central lymph node metastases, including logistic regression, support vector machine, random forest, decision tree, and adaptive boost. Results: Patients (n=229) were randomly divided into training (n=161) and validation (n=68) cohorts at a ratio of 7:3. Sixty-four patients exhibited central lymph node metastases. Logistic regression was the preferred algorithm to predict the occurrence of central lymph node metastases. The area under the curve, sensitivity, specificity, precision, recall, accuracy, and F1-score were 0.722, 0.761, 0.682, 0.833, 0.761, 0.735, and 0.795, respectively. Conclusions: Novel ultrasound radiomic machine learning models accurately predicted the occurrence of central lymph node metastases in patients with papillary thyroid carcinoma. The radiomic-based logistic regression model was the most effective and reliable preoperative method for the differential diagnosis of central lymph node metastases.