Optimizing C-TIRADS for sub-centimeter thyroid nodules using machine learning–derived feature importance

Background To optimize the diagnostic performance of the Chinese Thyroid Imaging Reporting and Data System (C-TIRADS) for sub-centimeter thyroid nodules by incorporating machine learning–derived feature importance. Methods This retrospective study included 741 patients in a primary cohort and 421 patients in an external validation cohort. SHapley Additive exPlanations (SHAP) were used to quantify the diagnostic contribution of six ultrasound features based on an XGBoost model. A modified C-TIRADS scoring system was developed by assigning greater weight to the most contributive feature while retaining original weights for other features. Diagnostic performance was evaluated using the area under the receiver operating characteristic curve (AUC), net reclassification improvement (NRI), and decision curve analysis (DCA). Results SHAP analysis identified vertical orientation as the most predictive feature for malignancy in sub-centimeter nodules. The modified scoring system significantly improved diagnostic performance in both the primary (AUC: 0.911 vs. 0.898, P < 0.001) and validation cohorts (AUC: 0.931 vs. 0.899, P < 0.001). NRI analysis further showed a substantial improvement in risk classifications, with NRI values of 0.406 in the primary and 0.471 in the validation cohort (both P < 0.001). DCA demonstrated greater net clinical benefit across wider threshold ranges in both cohorts. Additionally, malignancy rates exhibited a more rational stepwise increase from C-TIRADS 4A to 5, indicating improved risk stratification. Conclusion The SHAP-guided modified C-TIRADS scoring system enhances diagnostic accuracy and risk stratification for sub-centimeter thyroid nodules and may facilitate improved clinical decision-making in this challenging subset.