Development and Validation of a Predictive Model for Thyroid Nodule Malignancy Risk Based on Intralesional and Perilesional Ultrasound Radiomic Features

Abstract

Background: This study aimed to develop a multimodal integrated predictive model that combines intralesional and perilesional radiomic features with clinical characteristics, and to systematically evaluate its clinical utility in predicting malignancy among thyroid nodules (TNs) classified as C-TIRADS 4a/b.

Methods: A retrospective cohort of 147 patients with TNs who underwent thyroid ultrasound and subsequent ultrasound-guided fine-needle aspiration biopsy (US-FNAB) at Jiaxing Maternal and Child Health Care Hospital between August 2023 and September 2025 was enrolled. Participants were randomly allocated in a 7:3 ratio into training (n = 102) and validation (n = 45) cohorts. ITK-SNAP software was employed to delineate intralesional regions of interest (ROIs), with perilesional ROIs generated through a 2-mm outward expansion. The PyRadiomics module was utilized to extract radiomic features from horizontal intralesional, longitudinal intralesional, and integrated horizontal and longitudinal intra- and perilesional regions. Based on the selected features, three machine learning algorithms—logistic regression (LR), decision tree, and Light Gradient Boosting Machine—were implemented to develop radiomics models and clinical–radiomics combined models. Model performance was assessed using receiver operating characteristic curves, calibration curves, and decision curve analysis (DCA), with interpretability analysis conducted using the Shapley additive explanations framework.

Results: Multivariate LR analysis identified microcalcification, aspect ratio >1, and irregular shape as independent predictors of malignant TNs. The clinical–radiomics combined model demonstrated superior performance in the validation cohort, achieving an area under the curve of 0.866 (95% CI: 0.728–0.970) using the LR algorithm, with an accuracy of 84.4%, sensitivity of 91.3%, and specificity of 77.3%. DCA confirmed that the clinical–radiomics combined model provided the highest net clinical benefit across a wide range of threshold probabilities. SHAP analysis indicated that the combined intralesional–perilesional radiomics scores contributed most substantially to model predictions, highlighting the critical importance of radiomic features.

Conclusions: This study successfully developed a multimodal integrated predictive model combining intralesional and perilesional radiomic features with clinical characteristics. The model demonstrated excellent diagnostic performance in differentiating benign from malignant TNs classified as C-TIRADS 4a/b, providing clinicians with a convenient, non-invasive, and highly effective auxiliary diagnostic tool to support clinical decision-making.