Improving Non-Invasive Prediction of Thyroid Nodule Malignancy: A Machine Learning-Based Clinical Approach

Thyroid cancer (TC) is the most commonly diagnosed endocrine malignancy, with rising global incidence. Current diagnostic techniques, including ultrasound and fine-needle aspiration biopsy (FNAB), often yield inconclusive results, leading to unnecessary thyroidectomies for benign nodules. Improving preoperative risk stratification using non-invasive methods remains an important clinical challenge. This study aimed to develop machine learning (ML) models to enhance the classification of thyroid nodules (TNs) as malignant or benign based solely on selected ultrasonographic features.

Data from 5928 patients who underwent thyroidectomy at Wroclaw Medical University (2008-2023) were retrospectively analyzed. Five ultrasonographic features were included: hypoechogenicity, microcalcifications, shape, irregular margins, and vascularity. Five ML models - Random Forest, Logistic Regression, Multilayer Perceptron (MLP), Gradient Boosting Machines, and Decision Tree - were trained and evaluated. Model performance was assessed using accuracy, precision, recall, F1 score, specificity, and the area under the receiver operating characteristic curve (ROC-AUC). Feature importance was analyzed to determine the contribution of each variable.

Among the evaluated models, Random Forest achieved the highest overall performance, with an accuracy of 0.905, specificity of 0.939, ROC-AUC of 0.843, and recall of 0.616. Nodule vascularity was identified as the most influential predictor, followed by microcalcifications, irregular margins, and hypoechogenicity.

ML models based on a limited set of ultrasonographic features can effectively support the non-invasive identification of benign TNs, potentially reducing unnecessary surgical interventions. However, the modest recall underscores that the current approach is insufficient for reliable standalone malignancy detection. Incorporation of additional imaging parameters and cytological data would be necessary to enhance sensitivity and improve clinical applicability.