Explainable radiomics dictionary links thyroid US features to TI-RADS for AI-assisted nodule classification

An interpretable radiomics framework mapping quantitative US features to TI-RADS semantics achieved ROC-AUC 0.941 ± 0.004 for benign vs. malignant thyroid nodule classification across 5,542 multicenter nodules — with SHAP analysis confirming texture heterogeneity as the dominant…

• Retrospective multicenter study (5,542 nodules, three datasets); 107 radiomic features extracted, 27 feature-selection methods × 25 classifiers evaluated; best model (Select-From-Model + Extra-Trees) tested on a held-out 30% split — no independent external prospective validation reported.
• SHAP analysis aligned top predictive features (Gray Level Run Length Matrix non-uniformity, intensity dispersion, kurtosis) with established high-risk TI-RADS descriptors, providing a bridge between black-box outputs and clinical lexicon.
• Key limitation: retrospective design using 2D US images only; the dictionary and model have not been prospectively or externally validated, limiting immediate clinical generalizability.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.