Assessment of prognostic stratification in colorectal cancer liver metastases using preoperative CT habitat imaging

Abstract

Objective This study aims to explore the value of radiomic features derived from preoperative CT habitat imaging in prognostic stratification of patients with colorectal cancer liver metastases (CRLM), and to construct the predictive model incorporating clinical information for individualized risk assessment. Methods Retrospectively data from 196 patients with CRLM who underwent surgical resection were collected, including preoperative portal venous phase CT images and clinical information. First, tumor regions were segmented into habitat subregions using the K-means clustering algorithm to capture intra-tumoral heterogeneity. Subsequently, habitat radiomic features were extracted and dimensionality was reduced through Spearman correlation analysis followed by Least absolute shrinkage and selection operator regression, key features were then selected using the maximum relevance minimum redundancy method. Habitat models and combined clinical-habitat models were constructed based on Naive Bayes, multilayer perceptron (MLP), and XGBoost algorithms, respectively. Model performance was evaluated using receiver operating characteristic curves (ROC), decision curve analysis (DCA), and calibration curves. Results Compared with Naive Bayes and MLP, XGBoost demonstrated superior predictive performance in evaluating the prognosis of patients with CRLM. In both the training and validation cohorts, the combined clinical-habitat model achieved the best performance, with the AUC of 0.935 and 0.893, the sensitivity of 95.00% and 84.00%, and the specificity of 94.87% and 84.85%, respectively. Decision curve analysis showed that the predictive model provided a high net clinical benefit, and calibration curves confirmed good agreement between the predicted probabilities and the actual observed outcomes. The proposed model presents significant advantages in predictive accuracy compared with previous studies based only on traditional radiomics or clinical indicators. Conclusion Habitat imaging based on preoperative CT can effectively quantify intratumoral heterogeneity. The combined model constructed with machine learning algorithms exhibits excellent performance in prognostic stratification of patients with CRLM. This approach provides a non-invasive, high-precision novel tool for clinical prognostic evaluation, with important promotion value and promising application prospects.