Source-linked AI summary
Radiomics-Based Thyroid Nodule Classification with Label-Conditional Conformal Prediction
Laura Testa, Pierpaolo Brutti, Umberto Ferraro Petrillo
TL;DR
Ultrasound thyroid nodule assessment has interobserver variability, and many biopsies prove benign. The paper combines radiomic features, a GAM classifier, and label-conditional cross-conformal prediction; it reports competitive classification and uncertainty estimates, with some cases referred for clinical review.
Problem
Qualitative ultrasound risk stratification has interobserver variability, and at least half of biopsies prove benign.
Source: §1 · pp. 1–2Method
The framework compares six classifiers on PyRadiomics features and applies label-conditional cross-conformal prediction to quantify uncertainty.
Source: §1 · p. 2, §2.4 · p. 3, §2.5 · p. 4Results
At ϵ = 0.09, 63% of test cases received certain predictions, with conditional precision of 0.93–0.96 and conditional NPV of 0.71–0.77; the GAM achieved competitive classification performance.
Source: §3.3 · p. 6, §4 · p. 7Takeaways & Limitations
The framework supports high-confidence malignancy predictions while referring ambiguous cases to a clinician.
Source: §4 · p. 7Class imbalance affected specificity and NPV, and reliance on one dataset limits generalizability to other acquisition settings and patient populations.
Source: §4 · p. 7
Abstract
from arXiv · showhide
Thyroid nodule malignancy assessment relies on ultrasound imaging, yet current risk stratification systems suffer from substantial interobserver variability, contributing to a high rate of unnecessary biopsies. We propose a radiomics-based framework combining quantitative feature extraction via PyRadiomics with a Generalized Additive Model (GAM) classifier, selected among six candidates for its favorable trade-off between discriminative performance and interpretability. To complement point predictions with uncertainty estimates, Cross-Conformal Prediction with label-conditional guarantees is applied, flagging ambiguous cases for clinical review rather than forcing a binary decision. Experiments are conducted on TN5000, a publicly available ultrasound dataset of thyroid nodules. After Spearman correlation filtering, 36 features are retained and the GAM achieves an AUC of $0.89$, a sensitivity of $0.83$, and a specificity of $0.82$ on the validation set. At a significance level $ε=0.09$, selected to ensure empirical coverage guarantees on both classes, 63\% of nodules on the test set receive a certain (i.e., single-label) prediction, with conditional precision of $0.94$ for certain malignant and conditional Negative Predictive Value (NPV) of $0.73$ for certain benign cases, while the remaining 37\% are referred to the clinician.
1 Introduction
Thyroid nodule assessment depends on ultrasound, but visual risk stratification has interobserver variability, and at least half of biopsies prove benign. The paper proposes radiomics-based classification and label-conditional conformal prediction to address interpretability and uncertainty-quantification gaps.
1 Introduction
- At least half of thyroid biopsies prove benign, while qualitative ultrasound risk assessment introduces interobserver variability.These challenges motivate more reliable nodule risk assessment.Source: §1 · pp. 1–2
- Radiomics offers quantitative, operator-independent features, but existing approaches have limitations in interpretability and predictive uncertainty quantification.Prior thyroid classification studies reported competitive performance, with PyRadiomics offering standardized feature extraction.Source: §1 · p. 2
- The proposed framework compares classifiers on PyRadiomics features and uses label-conditional cross-conformal prediction to flag uncertain cases for review.The model is selected to balance discriminative performance and interpretability.Source: §1 · p. 2
2 Data and Methods
The methods use bounding-box regions from TN5000 for radiomic feature extraction and classification. They filter redundant features, compare six classifiers, and apply label-conditional cross-conformal prediction to produce class-specific coverage guarantees.
2.1 TN5000 Dataset
- TN5000 provides 5,000 ultrasound images with biopsy-confirmed labels and radiologist-annotated bounding boxes; images were cropped to these regions for feature extraction.The dataset split comprised 3,500 training, 500 validation, and 1,000 test images, with approximately 71% malignant and 29% benign training cases.Source: §2.1 · pp. 2–3
2.2 Radiomic Feature Extraction
- PyRadiomics extracted 102 features from seven feature families using 2D-mode ultrasound scans, intensity normalization, and a bin width of 25.The features include first-order statistics, shape-based 2D, GLCM, GLRLM, GLSZM, NGTDM, and GLDM.Source: §2.2 · p. 3
2.3 Feature Selection and Standardization
- A Spearman filter removed one feature from pairs with |ρ| > 0.90, leaving 36 features, which were standardized using training-set means and standard deviations.Features from GLSZM and GLDM were entirely removed; validation and test features used the training-set standardization parameters.Source: §2.3 · p. 3
2.4 Classification
- Six classifiers were trained with inverse-frequency class weights and compared on validation-set AUC, sensitivity, specificity, precision, NPV, and F1-score.The candidates were LR-LASSO, GAM, Random Forest, Extra Trees, XGBoost, and a two-hidden-layer MLP.Source: §2.4 · p. 3
2.5 Label-Conditional Cross-Conformal Prediction
- Label-conditional cross-conformal prediction uses K-fold calibration to guarantee coverage separately for each class and returns a prediction set for each test image.In the binary setting, the possible sets are certain benign, certain malignant, both labels, or empty.Source: §2.5 · pp. 3–4
3 Results
The GAM offered the strongest sensitivity-specificity trade-off among six classifiers, while label-conditional conformal prediction selected ε=0.09 and gave most test cases a single-label prediction. Feature-importance analysis also linked three top features to ACR TI-RADS descriptors.
3.1 Classifier Selection
- 0.83 vs 0.80 sensitivity, 0.82 vs 0.78 specificity, 0.93 vs 0.91 precision, and 0.62 vs 0.56 NPV favored GAM over LR-LASSO.All six models had AUC values of approximately 0.88–0.89; GAM was selected for its sensitivity-specificity trade-off and native interpretability.Source: §3.1 · p. 4, Table 1 · p. 5
3.2 GAM Feature Importance and Interpretation
- The top 10 features were ranked by mean AUC decrease after random permutation, with standard deviations reported across 50 repetitions.AUC was used as the performance metric for permutation importance given class imbalance.Source: Table 2 · p. 5, §3.2 · p. 5
- Mean, Entropy, and Elongation map respectively to echogenicity, composition, and taller-than-wide shape descriptors, with feature values associated with malignancy risk.The corresponding partial dependence plots are shown in Figure 2.Source: §3.2 · p. 5, Figure 2 · p. 6
3.3 CCCP Results
- At ε=0.09, 63% of test cases received certain predictions, with conditional precision of 0.93–0.96 and conditional NPV of 0.71–0.77.The remaining 37% received both labels and were left to the clinician.Source: Table 3 · p. 6, §3.3 · p. 6
- At ε=0.09, coverage was 0.91 for benign and 0.92 for malignant cases; at ε=0.10, benign coverage fell to 0.89, below target.The study selected ε=0.09 as the largest evaluated level achieving the target on both classes.Source: §3.3 · pp. 5–6
4 Conclusion
The paper concludes that a GAM paired with label-conditional cross-conformal prediction offers interpretable classification and uncertainty estimates, referring ambiguous cases to clinicians. Its findings are constrained by class imbalance, a single dataset, and bounding-box crops.
4 Conclusion
- The GAM and label-conditional cross-conformal prediction provided classification and uncertainty estimates, supporting high-confidence malignancy predictions while referring ambiguous cases to clinicians.GAM was selected among six candidates for its trade-off between discriminative performance and interpretability.Source: §4 · pp. 6–7
- Class imbalance affected specificity and NPV and prevented target benign-class coverage at ε=0.10; single-dataset evaluation and bounding-box crops also limit generalizability and feature accuracy.The authors note that crops may include surrounding thyroid tissue and introduce noise into radiomic features.Source: §4 · p. 7
- Planned directions include addressing class imbalance, validating across multiple datasets, adding nodule segmentation before feature extraction, and deepening clinical interpretation of radiomic features.Source: §4 · p. 7