What question did this study set out to answer?

Evaluate the ability of machine learning models to predict malignancy and metastasis in thyroid cancer.

February 11, 2026Open Access

Prediction of malignancy and metastasis of thyroid cancer by combined feature sets through advanced machine learning

Key Points

Evaluate the ability of machine learning models to predict malignancy and metastasis in thyroid cancer.
Examined 803 patients using ultrasound for thyroid nodules, including 369 benign and 433 malignant nodules.
Employed transfer learning to extract features from ultrasound images while incorporating clinical characteristics.
Analyzed multiple machine learning models such as gradient boosting and multilayer perceptron.
Measured model performance using area under the receiver operating characteristic curve (AUROC).
Achieved AUROC scores of 0.82 for thyroid cancer diagnosis and 0.78 for lymph node metastasis prediction.
Gradient boosting model optimized with 12 features showed the best performance in predicting malignancy.
Notable correlation of TI-RADS with both malignancy and lymph node metastasis.

Abstract

Predicting thyroid cancer (TC) and its accompanying metastasis remains challenging because of the complexity of the disease and its diverse pathological subtypes. The complexity and variability of TC necessitate innovative approaches that leverage machine learning (ML) and deep learning (DL) for enhanced accuracy in diagnosis and lymph node metastasis (LNM) prediction. 803 patients were examined using ultrasound for thyroid nodules, inculing 369 benign nodules, and 433 malignant nodules, of which 118 were LNM. All patients were diagnosed by pathologists based on preoperative and postoperative specimens. Transfer learning was used to extract the features of ultrasound two dimensional images. Using images and matched clinical characteristics, we incorporated two-dimensional imaging features and tabular clinical features to predict both malignancy and LNM. We considered several ML models, including a random forest classifier, gradient boost classifier, XGB classifier, Naïve Bayes, and multilayer perceptron. The area under the receiver operating characteristic curve was used to benchmark model performance. The AI model demonstrated notable diagnostic precision, achieving AUROC scores of 0.82 for TC diagnosis and 0.78 for LNM prediction. Crucially, the prediction model demonstrated a synergistic performance when both clinical characteristics and ultrasonographic images were combined, underscoring the complementary strengths of each data type in improving diagnostic accuracy. It should be noted that TI-RADS was significantly correlated with both malignancy and LNM. However, the optimized model was a gradient-boosting model with 12 features and a cross-validation AUROC of 0.82, indicating that imaging features can appreciably summarize the details of the ultrasonographic images for these patients. This study delves into the nuanced integration of imaging and clinical data through the advanced ML, encompassing both traditional ML techniques and DL, to enhance the diagnostic precision and therapeutic strategies for TC.

Read Full Paperexternally

Bookmark

View Full Paper

Cite This Study

Xie et al. (Mon,) studied this question.

synapsesocial.com/papers/698be001058ab1890a13bbeb https://doi.org/https://doi.org/10.1186/s12911-026-03372-w

Bookmark

View Full Paper