Lymph-Node Radiomics Outperforms Tumor-Based CT Analysis for Predicting Colorectal Cancer Spread

Colorectal cancer remains one of the most frequently diagnosed malignancies worldwide and a major contributor to cancer-related mortality, and whether the disease has spread to lymph nodes is among the most consequential questions a surgeon faces before an operation. Nodal status shapes the choice of neoadjuvant therapy, the extent of surgical resection, and the need for postoperative adjuvant treatment, yet conventional CT assessment still relies largely on morphologic criteria such as node size and shape, which remain insufficiently accurate for reliably identifying metastatic lymph nodes. A new proof-of-concept study now reports that a radiomics approach built directly on the lymph nodes themselves, rather than on the primary tumor, substantially outperformed the more conventional tumor-centered strategy in preoperative prediction of nodal metastasis.

The study, published in European Radiology Experimental, was a single-center retrospective analysis of patients with pathologically confirmed colorectal cancer who underwent preoperative contrast-enhanced CT and surgical resection at Tongren Hospital, Shanghai Jiao Tong University School of Medicine, between 2019 and 2023. After applying inclusion and exclusion criteria, 183 patients were enrolled and randomly divided at a 7:3 ratio, stratified by pathological lymph node status, into a training cohort of 128 patients and a validation cohort of 55. Baseline clinical characteristics showed no statistically significant differences between the two cohorts. The prediction task was defined at the patient level, asking whether metastasis was present anywhere in each patient, rather than attempting node-by-node classification, because individual nodes seen on preoperative CT could not be reliably matched to specific pathological nodes after resection.

The technical pipeline began with segmentation of the primary tumor and of visible regional lymph nodes on arterial-phase thin-section contrast-enhanced CT images, the sequence chosen because it was consistently available across patients and offered high spatial resolution for small-node delineation. Regional nodes were anatomically divided into three stations according to the Japanese Society for Cancer of the Colon and Rectum classification: D1 pericolic or perirectal nodes adjacent to the primary lesion, D2 intermediate nodes along the main feeding vessels, and D3 principal nodes at the root of the feeding vessels. Any visible regional node with a short-axis diameter of at least 3 mm was eligible for annotation. Segmentation used a segment anything model (SAM) as an interactive assistance tool to generate initial contours, followed by manual correction and expert review by a doctoral-level operator with five years of clinical experience and a senior expert with more than twenty years of experience in colorectal cancer; both were blinded to postoperative pathological nodal status.

Radiomics features were extracted with the PyRadiomics package from the segmented regions, covering first-order statistics, shape features, and texture matrices including gray-level co-occurrence, run-length, size-zone, dependence, and neighboring gray-tone difference features, computed on original and wavelet-filtered images with a fixed bin width of 25 gray-level units. For patients with multiple visible nodes in the same station, features were summarized with predefined aggregation statistics such as mean, maximum, median, and standard deviation, and pooled features across stations captured overall nodal phenotypes at the patient level. When a nodal station was absent in a patient, the corresponding features were treated as missing and imputed with median values. Feature selection, including Spearman correlation screening, recursive feature elimination, and LASSO regression with ten-fold cross-validation, was performed only in the training cohort, with normalization parameters estimated there and applied to the validation cohort to avoid information leakage.

Three radiomics strategies were then compared using five representative classifiers spanning different paradigms of tabular learning: decision tree, random forest, LightGBM, extreme gradient boosting, and the transformer-based TabPFN. In the validation cohort, the best-performing primary-tumor radiomics model achieved an area under the receiver operating characteristic curve (AUC) of 0.637, whereas the best-performing multi-station lymph-node radiomics model, based on the LightGBM classifier, reached an AUC of 0.864. Pairwise comparison using DeLong’s test showed the lymph-node model significantly outperformed the primary-tumor model (p = 0.0055). The combined strategy, which merged selected features from both tumor and nodes, achieved an AUC of 0.840, significantly better than the primary-tumor approach (p = 0.0210) but not significantly better than the lymph-node model alone (p = 0.2211). Because adding tumor features conferred no additional discriminatory benefit, the lymph-node model was retained as the final radiomics strategy.

The authors also benchmarked the approach against routine radiologic practice. Conventional CT-based nodal assessment, in which patients recorded as cN0 were classified as negative and those with cN1 or cN2 as positive, achieved an accuracy of 69.9%, sensitivity of 72.5%, specificity of 68.4%, and a binary cN assessment AUC of 0.704 for patient-level pathological nodal metastasis prediction. A clinical model built from preoperative variables including gender, cancer antigen 125, tumor size, and clinical T stage yielded a validation AUC of 0.634. Notably, univariate logistic regression of the clinical variables had not identified any statistically significant predictors, so the clinical subset was chosen based on prior literature, clinical relevance, and model parsimony rather than univariate significance alone. Both the final radiomics model and the clinical-radiomic model significantly outperformed the clinical model (p = 0.0199 and p = 0.0069, respectively).

Integrating the lymph-node radiomics probability with the selected clinical variables produced a clinical-radiomic model with a validation AUC of 0.879, the highest numerical value in the study, and a corresponding nomogram was derived for individualized preoperative risk estimation. However, the improvement over the radiomics model alone was not statistically significant (p = 0.2809), and at the fixed probability threshold of 0.5 the threshold-dependent metrics of the two models were identical in the validation cohort. Calibration curves in both cohorts showed acceptable agreement between predicted and observed probabilities, and decision curve analysis indicated the combined model yielded greater net benefit than treat-all and treat-none strategies across a range of threshold probabilities. The authors emphasized that the nomogram should be read as an interpretable visualization of the integrated model rather than evidence of a substantially superior predictor. The clinical model’s sensitivity of zero in the validation cohort at the fixed threshold also illustrated why threshold-dependent results were interpreted only descriptively, with AUC serving as the primary metric.

Interpretability and reproducibility analyses rounded out the evaluation. Shapley additive explanations (SHAP) analysis of the LightGBM-based lymph-node model visualized the relative contributions and nonlinear effects of the six selected radiomics features, whose full names, classes, filters, and aggregation statistics are listed in the supplementary materials. Five-fold cross-validation within the training cohort supported internal robustness of both the radiomics and clinical-radiomic models. In a bootstrap sensitivity analysis of 1,000 iterations in the validation cohort, the final radiomics model showed a bootstrap mean AUC of 0.865 with a 95% confidence interval of 0.744 to 0.960, a relatively wide interval the authors attributed to the limited validation-cohort size. Interobserver reproducibility, tested by independent re-segmentation of 30 randomly selected cases, showed that 1,769 of 2,007 extracted features (88.1%) had excellent agreement with intraclass correlation coefficients of at least 0.90, and 1,876 (93.5%) reached at least 0.75, with a mean ICC of 0.938. Five of the six final selected features showed good-to-excellent reproducibility, while one gray-level size-zone matrix feature showed limited reproducibility; model-predicted probabilities had limited absolute-agreement ICC but high Pearson and Spearman rank-order consistency between repeated segmentations.

The study’s limitations are substantial and were acknowledged explicitly. It was a single-center retrospective analysis with a moderate sample size, particularly in the validation cohort, and comparing multiple radiomics strategies and classifiers in a proof-of-concept design may introduce residual model-selection optimism even though feature selection and hyperparameter tuning were confined to the training cohort. All included patients underwent surgical resection, which may have introduced selection bias and limits applicability to patients with resectable colorectal cancer undergoing preoperative surgical evaluation. Because direct node-by-node pathological correlation was not performed, the model should not be interpreted as diagnosing metastasis in each segmented lymph node. CT spatial resolution cannot detect microscopic metastases in non-visible nodes, FDG-PET/CT was not routinely performed, systematic oncologic follow-up was not included so associations with recurrence-free or overall survival could not be evaluated, and uniform isotropic voxel resampling and formal scanner harmonization were not applied. The SAM-assisted segmentation workflow, while all masks required manual review, should be considered research-oriented rather than ready for routine clinical use.

The authors’ conclusion is measured: CT-based multi-station lymph-node radiomics achieved significantly higher validation performance than primary-tumor radiomics for patient-level preoperative prediction of lymph node metastasis in colorectal cancer, and direct characterization of regional lymph nodes across the D1, D2, and D3 stations may provide imaging information biologically closer to the metastatic process than tumor-centered features alone. The finding that a lymph-node-centered strategy preserved strong performance while remaining methodologically simpler than the combined approach is presented as methodologically and biologically informative. Nevertheless, the model is not ready for direct clinical implementation. Multicenter external validation with standardized CT acquisition and segmentation protocols, prospective evaluation, and fully nested validation are required before any clinical translation, and the proposed approach should be viewed as a potential adjunct to contrast-enhanced CT for nodal risk stratification rather than a replacement for other imaging modalities or pathological evaluation.

Subject of Research: CT-based multi-station lymph-node radiomics for preoperative prediction of lymph node metastasis in colorectal cancer

Article Title: CT-based multi-station lymph-node radiomics for preoperative prediction of lymph node metastasis in colorectal cancer: a proof-of-concept comparison with primary-tumor radiomics

Article References: Lin, Z., Wang, J., Chen, H., Bao, Q., Xu, M., Yang, S., Zhang, H., Yang, G.-Z., & Sun, P. (2026). CT-based multi-station lymph-node radiomics for preoperative prediction of lymph node metastasis in colorectal cancer: a proof-of-concept comparison with primary-tumor radiomics. European Radiology Experimental, 10(1), Article 153. https://doi.org/10.1186/s41747-026-00814-y

Image Credits: AI Generated

DOI: 10.1186/s41747-026-00814-y

Keywords: colorectal cancer, radiomics, lymph node metastasis, computed tomography, machine learning, LightGBM, clinical-radiomic nomogram, preoperative prediction, proof-of-concept study, single-center retrospective study, SAM-assisted segmentation, DeLong test