AI Reads Multi-Phase CT Scans to Tell Benign Thyroid Nodules From Cancer

Thyroid nodules are extraordinarily common: ultrasound scans pick them up in roughly 65 percent of the general population, and while the vast majority are harmless, 5 to 15 percent turn out to be malignant. Deciding which nodules need surgery and which can be safely watched is one of the most consequential judgment calls in endocrine medicine. A new multicenter study published in Frontiers in Endocrinology offers a strikingly quantitative answer to that problem, showing that artificial intelligence analysis of routine contrast-enhanced CT scans can distinguish cancerous thyroid nodules from benign ones with an accuracy that significantly exceeds both clinical judgment alone and single-modality radiomics models.

The research team, led by investigators at Wenzhou Medical University Affiliated Quzhou Hospital and Ningbo No. 2 Hospital in China, built what radiologists call a radiomics pipeline. Radiomics is a technique that extracts hundreds of quantitative features from medical images—grayscale statistics, texture patterns, shape descriptors, and measures of spatial heterogeneity—that are invisible to the human eye. Where a radiologist sees a nodule’s outline and density, the algorithm sees a numerical fingerprint of the tumor’s internal architecture, capturing information about cellularity, fibrosis, and microvascular structure that correlates with underlying biology.

To build and test their models, the researchers retrospectively enrolled patients with surgically confirmed thyroid nodules who had undergone preoperative contrast-enhanced thyroid CT at two hospitals. At the first center, 725 patients with 417 benign and 308 malignant nodules were randomly split 8:2 into a training cohort of 580 patients and an internal validation cohort of 145. A completely independent group of 208 patients from the second hospital—58 benign and 150 malignant nodules—served as the external validation cohort, the gold standard for testing whether a model generalizes beyond the data it was trained on. Only solitary nodules at least 5 millimeters in diameter were included, and patients with prior thyroid surgery, radiotherapy, chemotherapy, or biopsy before CT were excluded to avoid contaminating the imaging signature.

The team extracted radiomics features from three CT phases: unenhanced images, arterial-phase images acquired roughly 20 to 30 seconds after contrast injection, and venous-phase images acquired 35 to 60 seconds after injection. All images were resampled to a uniform 1-millimeter isotropic voxel size with standardized window settings before feature extraction, a preprocessing step designed to minimize differences between the two centers’ scanners—one a 320-detector-row Toshiba unit, the other a 64-slice Siemens system. In total, 1,834 features were pulled from each phase, yielding 5,502 candidate features for the multiphase model. Through a multi-step selection process involving statistical filtering, Pearson correlation pruning, minimum redundancy maximum relevance ranking, and LASSO regression, the researchers distilled this down to 25, 19, and 35 features for the unenhanced, arterial, and venous phases respectively, and 35 features for the final multiphase signature.

Four categories of models were constructed and compared: a clinical model based on patient characteristics and conventional CT morphology, single-phase radiomics models for each of the three CT phases, a multiphase radiomics model combining features from all phases, and a combined clinical-radiomics model integrating everything. The classifier was a multilayer perceptron—a type of neural network—with two hidden layers of 128 and 64 neurons, ReLU activation, dropout regularization, and the Adam optimizer. Crucially, all feature selection and model training were performed strictly within the training cohort, with the external validation cohort held entirely apart to prevent information leakage.

The results were revealing. The multiphase radiomics model achieved an area under the ROC curve (AUC) of 0.919 in the training cohort, 0.836 in internal validation, and 0.787 in external validation. Among single-phase models, the venous-phase model proved the most robust across centers, achieving an external validation AUC of 0.815—actually outperforming the multiphase model on unseen data from the second hospital. The researchers attribute this to the venous phase more reliably capturing differences in stromal enhancement, contrast retention, and fibrotic content within lesions, while avoiding the transient perfusion fluctuations that plague arterial-phase imaging. Previous studies have disagreed about which phase is optimal; this multicenter design, with standardized preprocessing, suggests venous-phase features may be the most transportable.

The clinical model, built from seven independent predictors identified by multivariate logistic regression—nodule diameter, patient age, total thyroxine, free thyroxine, bite sign, irregular shape, and range reduction—achieved an external validation AUC of 0.800. These morphological signs have clear pathological underpinnings: irregular shape reflects infiltrative growth, the bite sign indicates capsular contact or invasion, and range reduction describes a decrease in density contrast between the nodule and surrounding thyroid tissue on enhanced images, a hallmark of abnormal vascularity. Interestingly, microcalcification, often considered a classic red flag, did not reach statistical significance in this cohort, possibly due to limited sample size and the small number of participating centers.

When clinical and morphological features were combined with multiphase radiomics, the integrated model achieved the best overall performance: an external validation AUC of 0.865, with 74.0 percent sensitivity and 84.5 percent specificity. This significantly outperformed both the clinical model alone (AUC 0.800) and the multiphase radiomics model alone (AUC 0.787), as confirmed by DeLong’s statistical test. Calibration curve analysis showed the combined model had the lowest Brier score, meaning its predicted probabilities tracked actual malignancy risk most closely, and decision curve analysis demonstrated greater clinical net benefit across most threshold probabilities—the metric that matters most when deciding whether a prediction is useful for real patient management.

Perhaps the most compelling aspect of the study is its commitment to interpretability. Rather than presenting a black box, the researchers applied SHAP (SHapley Additive exPlanations) analysis to quantify how much each feature contributed to every individual prediction. At the population level, a beeswarm plot revealed that features reflecting local grayscale heterogeneity and inter-voxel textural correlation—measures of intralesional structural complexity—drove malignant predictions, while benign nodules were characterized by first-order statistics and textural uniformity. At the individual level, two representative patients were examined in detail: one with a pathologically confirmed benign nodule received an accurate low-risk prediction with a SHAP value of 0.051, while another with proven malignancy was correctly flagged as high risk with a SHAP value of 0.961. Both predictions matched postoperative histopathology, with clear directional contributions from specific features.

The authors are careful to position this tool correctly: it does not replace ultrasound and the ACR TI-RADS classification system, which remain the foundation of thyroid nodule evaluation. Instead, CT radiomics serves as a valuable adjunct for the growing population of patients whose nodules are discovered incidentally on CT scans performed for other reasons, for those with indeterminate ultrasound findings, or for patients needing detailed assessment of the nodule’s relationship to the trachea, vessels, and cervical lymph nodes before surgery. By providing an objective, noninvasive risk estimate, the combined model could help reduce unnecessary fine-needle aspirations and surgeries while ensuring that genuinely malignant nodules are not overlooked. The researchers acknowledge limitations—retrospective design, inclusion of only solitary nodules, manual segmentation that is labor-intensive, and the absence of head-to-head comparison with ultrasound-based radiomics—but they argue that future work incorporating automated segmentation, deep learning, and molecular biomarkers could push performance even further. For now, the study stands as one of the most rigorously validated demonstrations that the humble CT scan, long considered a secondary tool in thyroid imaging, can be transformed into a quantitative diagnostic instrument through the lens of machine learning.

Subject of Research: Multiphase CT radiomics for preoperative differentiation of benign and malignant thyroid nodules

Article Title: Computed tomography multi-phase radiomics models for differentiating benign and malignant thyroid nodules

Article References: Li, N., Tong, J., Zhang, G., Xu, S., Huang, W., Han, X., Zhang, J., & Zhu, X. (2026). Computed tomography multi-phase radiomics models for differentiating benign and malignant thyroid nodules. Frontiers in Endocrinology, 17, Article 1928263. https://doi.org/10.3389/fendo.2026.1928263

Image Credits: AI Generated

DOI: 10.3389/fendo.2026.1928263

Keywords: thyroid nodules, radiomics, computed tomography, artificial intelligence, cancer diagnosis, machine learning, medical imaging, SHAP, external validation, preoperative risk stratification, venous phase, clinical model