MRI habitat radiomics maps tumor diversity to predict lymph node spread in early-stage cervical cancer

Researchers have built an interpretable imaging framework that clusters cervical tumors into distinct MRI-defined subregions and uses those patterns to predict whether cancer has spread to nearby lymph nodes before surgery. In a study published in Medical Oncology, the multiparametric MRI habitat radiomics approach reached a test area under the curve of 0.841 on its own and 0.890 when combined with clinical and radiological variables, offering a preoperative alternative to methods that average out a tumor’s internal diversity.

Preoperative assessment of lymph node metastasis informs treatment planning in early-stage cervical cancer, yet traditional radiomics compresses each tumor into a single set of quantitative features derived from the whole lesion. That wholesale averaging can obscure intratumoral heterogeneity, the fact that different parts of the same tumor may behave differently biologically. The new framework instead divides the tumor into habitats, spatially coherent clusters of voxels that share similar signal characteristics across three MRI sequences.

The study was organized in two parts. Part I included 148 patients for training and 63 patients for testing, all of whom underwent radical surgery with lymphadenectomy so that imaging predictions could be checked against surgical pathology. Part II comprised 14 patients who provided spatially matched histopathology, allowing the investigators to link each MRI-defined habitat to tissue sampled from the corresponding tumor region.

To construct the habitats, intratumoral voxels were clustered by K-means using signals from T2-weighted imaging, diffusion-weighted imaging, and contrast-enhanced T1-weighted imaging. The number of habitats was not fixed in advance; it was selected using validation AUC, and a four-habitat solution proved best. From these subregions the team derived 14 features based on DWI and CE-T1WI signals, which fed a habitat radiomics score. Habitat radiomics and traditional radiomics scores were each integrated with clinical-radiological predictors through logistic regression, and model performance was compared with DeLong’s and chi-squared tests.

Lymph node metastasis was present in 27.1 percent of the 225 patients, or 61 cases, giving the models a clinically meaningful prevalence to detect. On the independent test set, the clinical model alone achieved an AUC of 0.804 and traditional radiomics reached 0.784, while habitat radiomics combined with clinical variables reached 0.890. Adding traditional radiomics to the habitat and clinical model produced the highest AUC of 0.902 and an accuracy of 87.30 percent, but the difference from the habitat-plus-clinical model was not statistically significant, with a P value of 0.277. The authors therefore describe the habitat-plus-clinical model as a parsimonious candidate for further evaluation, since it achieves comparable performance without the extra radiomics component.

Interpretability was addressed with Shapley additive explanations, a technique that quantifies each variable’s contribution to the model’s predictions. SHAP ranked the habitat radiomics score above clinical lymph-node status, tumor diameter, and the traditional radiomics score, indicating that the subregion-based features carried the greatest predictive weight in the combined model.

The exploratory histopathologic analysis in Part II gave the habitats a biological context. Subregion-matched tissue samples were quantified for Ki-67, a proliferation marker; CD3, a T-cell marker; CD31, an endothelial marker of microvessel density; and the tumor-stroma ratio. Habitat 3 showed the highest Ki-67 expression together with the lowest tumor-stroma ratio, CD3, and CD31, marking it as the most proliferative and least immune- and vessel-rich and stroma-dense subregion. Habitat 1 displayed the highest tumor-stroma ratio and the lowest Ki-67, while habitat 2 showed higher CD31 than habitat 4. The authors note that this spatially matched exploratory histopathology suggests possible regional tissue differences across the MRI-defined habitats rather than confirming a definitive biological mechanism.

Several caveats frame the findings. The test cohort contained 63 patients, the histopathologic correlation rested on only 14 cases, and the reported AUCs and accuracy come from a single study population. The paper’s conclusion is correspondingly measured: MRI-based habitat radiomics may support preoperative lymph node assessment in early-stage cervical cancer, with the combined habitat and clinical model identified as a candidate that requires further evaluation before clinical use. Because all patients underwent radical surgery and lymphadenectomy, the predictions could be validated against actual nodal status, but external validation in other populations remains a necessary next step.

What distinguishes the work from conventional radiomics studies is the pairing of a habitat-based predictive model with an interpretable structure and a spatially anchored pathological check. Rather than treating the tumor as a homogeneous object, the framework preserves its internal architecture in the feature set, and the SHAP analysis makes the relative contributions of imaging and clinical variables explicit. For patients with early-stage cervical cancer, the practical promise is a more informative preoperative picture of nodal involvement, drawn from MRI sequences that are already part of routine staging, with the habitat-derived score carrying much of the predictive signal.

Subject of Research: An interpretable multiparametric MRI habitat radiomics framework for preoperative prediction of lymph node metastasis in early-stage cervical cancer, with exploratory spatially matched histopathology.

Article Title: Deciphering intratumoral heterogeneity: An interpretable radiopathologic framework for lymph node staging in early-stage cervical cancer

Article References: Chen, J., Shi, Y., Song, Y., Liu, G., Zhou, Y., Wei, X., Zhu, M., Luo, Y., Zhou, B., Hu, Z., Song, M., & Wu, G. (2026). Deciphering intratumoral heterogeneity: An interpretable radiopathologic framework for lymph node staging in early-stage cervical cancer. Medical Oncology, 43(12), Article 369. https://doi.org/10.1007/s12032-026-03492-3

Image Credits: AI Generated

DOI: 10.1007/s12032-026-03492-3

Keywords: cervical cancer, lymph node metastasis, habitat radiomics, multiparametric MRI, intratumoral heterogeneity, SHAP interpretability, Ki-67, tumor-stroma ratio, machine learning, radiopathology, DWI, CE-T1WI