CT-based habitat radiomics flags heavy nodal spread in thyroid cancer, with caveats

A retrospective study of 119 patients with papillary thyroid carcinoma reports that a machine-learning model built from imaging-defined subregions of the thyroid gland on contrast-enhanced CT distinguished patients with a heavy burden of central compartment lymph node metastasis from those with a lighter burden. In an independent test cohort of 48 patients from a second institution, the model achieved an area under the receiver operating characteristic curve of 0.913, numerically higher than a clinical model built from nine routine preoperative variables, which reached 0.731. The authors themselves, however, flag substantial overfitting and note that the advantage over a simple model using only age and sex was not statistically significant.

The study, published in the Egyptian Journal of Radiology and Nuclear Medicine, addressed a practical problem in thyroid cancer care. Papillary thyroid carcinoma is the most common endocrine malignancy, and cervical lymph node metastasis is present in a substantial proportion of patients at diagnosis. Nodal spread is associated with an increased risk of locoregional recurrence and with more complex subsequent management. Identifying a heavy central compartment nodal burden before surgery could prompt more complete neck dissection or inform decisions about adjuvant radioiodine therapy, but conventional imaging and clinical variables discriminate poorly.

The research team defined high-burden disease as three or more pathologically positive central compartment lymph nodes, with the low-burden reference group comprising patients with zero to two positive nodes. This threshold was an operational choice made for the analysis. The authors point out that it does not restate the 2015 American Thyroid Association intermediate-risk criterion of more than five involved nodes; applied to this cohort, that criterion would have placed 93 of the 119 patients, or 78.2 percent, in the low-burden group, leaving too few high-burden patients for a meaningful comparison. Three or more positive nodes classified 69 of the 119 patients, or 58.0 percent, and the median number of retrieved central compartment nodes was nine, with a range of seven to thirteen.

Methodologically, the study was retrospective and conducted at two independent institutions. Seventy-one patients treated at Centre A formed the training cohort and 48 patients treated at Centre B formed the independent-centre test cohort; no patient appeared in both. All participants had histologically confirmed papillary thyroid carcinoma, had undergone thyroid surgery with central compartment neck dissection, and had preoperative contrast-enhanced neck CT with complete nodal records. Baseline characteristics were similar between cohorts: mean age was 43.82 years in the training cohort and 45.56 years in the test cohort, male sex accounted for roughly 35 percent in both, and high-burden nodal disease was present in 60.6 percent and 54.2 percent of patients respectively.

The imaging pipeline departed from the usual radiomic practice of outlining the primary tumour. Because papillary thyroid carcinoma nodules are frequently smaller than one centimetre and poorly demarcated on CT, the entire thyroid gland, including tumour and surrounding parenchyma, was delineated as a single region of interest by two radiologists with more than eight years of experience in head and neck imaging. The images then underwent super-resolution reconstruction that tripled the in-plane sampling, increasing the in-plane pixel count ninefold while leaving slice thickness unchanged. The authors are explicit that this step interpolates the CT volume onto a finer grid and does not recover spatial resolution beyond that of the original acquisition, so submillimetre texture features should be read as interpolated quantities rather than direct measurements.

Within the delineated gland, the team applied habitat analysis, an approach that partitions a tissue volume into subregions with distinct imaging phenotypes rather than treating it as a single lesion. Nineteen standardised imaging features were extracted voxelwise, and unsupervised k-means clustering of the pooled voxels identified four habitat subregions as the optimal configuration, selected by the Calinski-Harabasz criterion among solutions from two to seven clusters. The four subregions differed systematically in attenuation, with mean Hounsfield unit values of approximately 107, 156, 166 and 129, and their volumetric composition was consistent across patients. From these subregions, 428 radiomic features were extracted per patient. Feature selection, model fitting and threshold derivation were performed in the training cohort only, with the notable exception that the habitat clustering itself was fitted on pooled voxels from the whole cohort, which the authors report as a limitation.

After a two-step filtering and LASSO selection process retained 17 features, seven classifiers were trained and compared, and XGBoost showed the highest performance. In the independent test cohort, the habitat-only model achieved an AUC of 0.913 with a 95 percent confidence interval of 0.831 to 0.994. At a threshold derived in the training cohort, sensitivity was 0.808 and specificity 0.955. The nine-variable clinical comparator, which included thyroid antibodies, thyroglobulin, age, thyroid function measures, Hashimoto’s thyroiditis and sex, reached an AUC of 0.731. The difference of 0.182 was statistically significant in a DeLong test with a p value of 0.040, but the authors stress that all pairwise comparisons were exploratory and unadjusted for multiplicity and should not be read as confirmatory evidence of superiority.

Two caveats qualify the headline result. First, when the clinical comparator was reduced to age and sex alone, a sensitivity analysis motivated by the fact that age showed the strongest univariable association with nodal burden, the habitat model’s advantage shrank to 0.136 in AUC and was no longer statistically significant, with a p value of 0.105. Younger patients with high-burden disease were consistently seen in both cohorts, with mean ages around 40 years versus 50 to 52 years for low-burden patients. Second, overfitting was substantial: with 428 features available for 71 training patients, the apparent training AUC reached 1.000. A repeated nested cross-validation sensitivity analysis under an alternative pipeline gave an estimate of 0.676, implying an optimism of roughly 0.32, so the test cohort estimate of 0.913 is unlikely to be reproduced exactly in a new cohort.

Additional analyses tempered other aspects of the framework. On original-resolution images without super-resolution reconstruction, the habitat model achieved a test AUC of 0.867, and none of the two-by-two comparisons between habitat partitioning and whole-gland radiomics, or between reconstructed and original images, reached statistical significance. Only 8 of the 17 features retained by the LASSO model were stable across reconstructions, so the signature should not be described as resolution-invariant. Calibration was also unstable, with calibration slopes well above the ideal value of one and wide confidence intervals in the small test cohort. The authors note that a non-significant Hosmer-Lemeshow test does not by itself establish good calibration.

The authors interpret the four subregions cautiously. Because the clustering is unsupervised and label-free, the subregions cannot be assigned a biological identity on the basis of the present data, and no histopathological or molecular correlation was performed. The whole-gland region of interest necessarily includes normal parenchyma, Hashimoto-related inflammatory infiltrates and colloid in addition to any malignant focus, so the extracted features describe intra-glandular rather than tumour-specific heterogeneity. The team concludes that the approach should be regarded as preliminary and hypothesis-generating, and that prospective multicentre studies with tumour-specific segmentation, larger samples and histopathological correlation are required before any clinical use can be considered.

Subject of Research: Habitat radiomics on contrast-enhanced CT for predicting high-burden central compartment lymph node metastasis in papillary thyroid carcinoma

Article Title: Predicting high-burden central compartment lymph node metastasis (≥ 3 positive nodes) in papillary thyroid carcinoma: a habitat radiomic framework based on reconstructed contrast-enhanced CT

Article References: Sun, D., Li, H., Zhan, Y., & Zhang, Z. (2026). Predicting high-burden central compartment lymph node metastasis (≥ 3 positive nodes) in papillary thyroid carcinoma: a habitat radiomic framework based on reconstructed contrast-enhanced CT. Egyptian Journal of Radiology and Nuclear Medicine, 57(1), Article 225. https://doi.org/10.1186/s43055-026-01894-x

Image Credits: AI Generated

DOI: 10.1186/s43055-026-01894-x

Keywords: papillary thyroid carcinoma, lymph node metastasis, radiomics, habitat analysis, contrast-enhanced CT, super-resolution reconstruction, XGBoost, central compartment, overfitting, retrospective study, machine learning, thyroid gland