[Paper Review] A Clinically Anchored Radiomics Dictionary for Explainable TI-RADS-Based Thyroid Nodule Classification in Ultrasound; Dictionary Version TU1.0

Mohammad Salmanpour, Shahram Taeb+6 more|arXiv (Cornell University)|Mar 24, 2026

Thyroid Cancer Diagnosis and Treatment0 citations

TL;DR

The paper presents a clinically anchored radiomics framework linking US radiomic features to TI-RADS semantics, achieving explainable thyroid nodule classification with a high ROC-AUC on multicenter data.

ABSTRACT

Artificial intelligence based radiomics models for thyroid ultrasound (US) often achieve strong diagnostic performance but remain difficult to interpret, limiting clinical trust and adoption. We developed and validated an interpretable radiomic feature (RF) framework for thyroid nodule classification by linking quantitative US features to the Thyroid Imaging Reporting and Data System (TI-RADS) semantic lexicon through a clinically grounded radiomics dictionary. The dictionary mapped TI-RADS categories, including composition, echogenicity, shape, margin, and echogenic foci, to Image Biomarker Standardization Initiative compliant RFs extracted from two-dimensional US images. Relationships were defined through expert consensus and examined using Shapley Additive Explanations (SHAP). Three multicenter datasets were combined, yielding 5,542 nodules. A total of 107 RFs were extracted using PyRadiomics and normalized with min-max scaling. For benign versus malignant classification, 27 feature selection methods were paired with 25 classifiers and evaluated using stratified five-fold cross-validation on 70% of the data, followed by testing on the remaining 30%. Robust model selection used a stability-aware composite score combining mean performance and variability across balanced accuracy, precision, recall, F1-score, and ROC-AUC. The proposed dictionary enabled direct interpretation of radiomic signatures in TI-RADS terms. The best model, Select-From-Model based on logistic regression with Extra-Trees, achieved a test ROC-AUC of 0.941 +/- 0.005. SHAP analysis showed that texture heterogeneity was the dominant malignancy signal, with gray level run length matrix non-uniformity, intensity dispersion, and kurtosis aligning with high-risk TI-RADS descriptors. These findings support transparent and clinically meaningful thyroid nodule risk stratification from US.

Motivation & Objective

• Bridge quantitative radiomics features with TI-RADS semantic descriptors to improve interpretability and clinical trust in thyroid US classification.
• Develop and validate a radiomics dictionary that maps TI-RADS categories to Image Biomarker Standardization Initiative (IBSI) compliant radiomic features.
• Evaluate robust model selection across multiple feature selectors and classifiers on multicenter data.
• Demonstrate interpretability via SHAP analysis linking radiomic signatures to TI-RADS risk descriptors.

Proposed method

• Extract 107 radiomic features from 2D US images using PyRadiomics and normalize with min-max scaling.
• Construct a dictionary mapping TI-RADS categories (composition, echogenicity, shape, margin, echogenic foci) to RFs.
• Combine three multicenter datasets to total 5,542 nodules for training and testing.
• Evaluate 27 feature selection methods with 25 classifiers using stratified five-fold cross-validation (70% training, 30% testing).
• Select robust models using a stability-aware composite score across balanced accuracy, precision, recall, F1, and ROC-AUC.
• Apply SHAP to interpret feature contributions and relate them to TI-RADS descriptors.

Experimental results

Research questions

• RQ1 Can a radiomics dictionary grounded in TI-RADS semantics yield interpretable thyroid nodule classifications in ultrasound?
• RQ2 Which radiomic features and TI-RADS descriptors most strongly indicate malignancy according to SHAP?
• RQ3 What is the performance of robust, dictionary-guided models across multicenter data for benign vs malignant thyroid nodules?

Key findings

• Best model (Select-From-Model with logistic regression and Extra-Trees) achieved test ROC-AUC 0.941 ± 0.005.
• A total of 107 radiomic features were extracted and used across 27 feature selectors and 25 classifiers.
• SHAP identified texture heterogeneity as the dominant malignancy signal; features like gray level run length matrix non-uniformity, intensity dispersion, and kurtosis aligned with high-risk TI-RADS descriptors.
• The dictionary enables direct interpretation of radiomic signatures in TI-RADS terms, supporting transparent thyroid nodule risk stratification.