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

Preoperative assessment of lymph node metastasis (LNM) informs treatment planning in early-stage cervical cancer (CC). Traditional radiomics (TR) may obscure intratumoral heterogeneity. We developed multiparametric MRI habitat radiomics (HR) for LNM prediction and explored its histopathologic context. Patients underwent radical surgery and lymphadenectomy. Part I comprised 148 training and 63 test patients; Part II comprised 14 for spatially matched histopathology. Intratumoral voxels were clustered into habitats by K-means using T2-weighted imaging (T2WI), diffusion-weighted imaging (DWI), and contrast-enhanced T1-weighted imaging (CE-T1WI) signals. The number of habitats was selected using validation AUC. HR/TR scores and clinical-radiological predictors were integrated by logistic regression. AUC and accuracy were compared by DeLong's and χ² tests. Shapley additive explanations (SHAP) assessed contributions. Subregion-matched histopathology quantified Ki-67, CD3, CD31, and tumor-stroma ratio (TSR). LNM prevalence was 27.1% (61/225). The four-habitat HR model achieved a test AUC of 0.841 with 14 DWI/CE-T1WI-derived features. Test AUCs for clinical model (CL), TR, and HR + CL were 0.804, 0.784, and 0.890. HR + TR+CL achieved the highest AUC (0.902) and accuracy (87.30%), without a significant AUC difference from HR + CL (P = 0.277). SHAP ranked HR score above clinical lymph-node status, tumor diameter, and TR score. HR-3 exhibited highest Ki67 and lowest TSR, CD3, and CD31; HR-1 had highest TSR and lowest Ki67; HR-2 showed higher CD31 than HR-4. MRI-based HR may support preoperative LNM assessment in early-stage CC, with HR + CL representing a parsimonious candidate model for further evaluation. Spatially matched exploratory histopathology suggests possible regional tissue differences across MRI-defined habitats.