Abstract Background Thyroid nodules with indeterminate cytology pose a diagnostic challenge, often leading to unnecessary thyroidectomies. Radiomics, by quantitatively analysing ultrasound images, may enhance diagnostic objectivity and improve preoperative decision-making. Method A retrospective single-center study was conducted at the University Hospital of Cagliari, including 69 patients with Bethesda IV (B4) thyroid nodules who underwent thyroid surgery. Radiomic features were extracted from ultrasound images using LIFEx software and processed through Principal Component Analysis (PCA) for dimensionality reduction. Three supervised machine learning algorithms—K-Nearest Neighbors (KNN), Random Forest (RF), and Extreme Gradient Boosting (XGBoost)—were compared using k-fold and Leave-One-Out Cross Validation. Results KNN yielded poor diagnostic performance (accuracy 61.9%, high false negatives), proving unsuitable for clinical application. Conversely, the Random Forest model achieved excellent performance, with accuracy up to 100% in training and 88–93% in validation (F1-score 0.89–0.94), while drastically reducing false negatives. XGBoost demonstrated comparable or superior results, achieving a mean accuracy of 93% and an AUC-ROC of 0.97, with over 90% sensitivity in validation. Both RF and XGBoost exhibited strong stability and generalisation, confirming their reliability as predictive tools for distinguishing benign from malignant nodules. Conclusion Radiomic-based machine learning models, particularly Random Forest and XGBoost, showed high diagnostic accuracy and sensitivity in differentiating B4 nodules. These results support the integration of radiomics into clinical workflows as a promising adjunct to conventional ultrasound, potentially reducing unnecessary thyroid surgeries and improving patient management.
Medas et al. (Fri,) studied this question.