A model integrating CEA and KRAS mutation statuses and computed tomography tumor-habitat radiomics predicts synchronous liver metastases in patients with rectal cancer: a dual-center study

Objective To investigate the diagnostic performance of a computed tomography (CT)-based habitat radiomics model for discriminating between patients with rectal cancer (RC) with and without synchronous liver metastases (SLM). Methods This retrospective study enrolled two independent patient cohorts. Cohort 1 (internal cohort) included 198 patients with pathologically confirmed RC with complete CT data from April 2022 to June 2025, comprising 46 SLM-positive and 152 SLM-negative cases. This cohort was divided randomly in a 7:3 ratio into a training set (140 patients: 33 SLM-positive, 107 SLM-negative) and an internal validation set (58 patients: 13 SLM-positive, 45 SLM-negative). Cohort 2 served as an external validation set, consisting of 55 patients with RC (18 SLM-positive, 37 SLM-negative). Regions of interest (ROIs) were delineated manually slice-by-slice on arterial-phase thin-slice CT images. Raw radiomics features were extracted from the ROIs and K-means clustering was implemented to segment tumor sub-habitats, followed by feature extraction for each sub-habitat. Feature-dimensionality reduction was performed on the training set using minimum redundancy maximum relevance and least absolute shrinkage and selection operator algorithms. Logistic regression classifiers were constructed to develop a conventional radiomics model and a habitat radiomics model, respectively. Model performance and clinical utility were quantified using the area under the receiver operating characteristic curve (AUC), calibration curves, and decision curve analysis (DCA). Results Univariate and multivariate logistic regression identified serum CEA and KRAS mutations as independent risk factors for SLM. These were integrated to build a clinical baseline model with AUC values of 0.730 (training set), 0.724 (internal validation set), and 0.701 (external validation set). After rigorous feature screening, 13 conventional radiomics features and 13 habitat radiomics features were retained for model construction. The conventional radiomics model yielded AUCs of 0.785 (training), 0.764 (internal validation), and 0.782 (external validation), while the habitat radiomics model achieved higher AUCs, with AUCs of 0.862, 0.826, and 0.826, respectively. The combined model integrating independent clinical risk factors and habitat radiomics score attained the numerically highest AUCs, with AUCs of 0.883 (training), 0.846 (internal validation), and 0.851 (external validation), outperforming the clinical, conventional radiomics, and standalone habitat models with robust cross-cohort stability. Calibration curves demonstrated favorable predictive consistency of the combined model, while DCA revealed sustained clinical net benefit across broad risk-threshold ranges. Conclusion A CT-based habitat radiomics model could effectively identify patients with RC with SLM, with performance numerically better than that of conventional whole-tumor radiomics. Integrating clinical biomarkers with habitat radiomics further increased the predictive accuracy, offering a novel non-invasive preoperative risk-stratification strategy for RC.