Integrated radiomics and machine learning approach for ras mutation status prediction in colorectal liver metastases

RAS mutational status is a critical prognostic biomarker in colorectal liver metastases (CRLM), traditionally assessed via tissue biopsy. This study evaluates the potential of radiomic features extracted from CT and MRI to non-invasively predict RAS mutations using machine learning algorithms.

In this study, 77 CRLM metastases (mean size 34.9; range 17-56 mm) with known RAS mutational status were analyzed. Radiomic features were extracted from hepatobiliary-phase MRI and portal venous-phase CT. After removing highly correlated features (Pearson |r| > 0.7) and applying z-score normalization, LASSO logistic regression with repeated tenfold cross-validation was used for feature selection. A total of 41 predictive features were identified. The dataset was split into training (70%) and testing (30%), ensuring that all lesions from a given patient were assigned exclusively to either the training or testing set. To address class imbalance in the training data, the Random Over Sampling Examples (ROSE) algorithm was applied exclusively to the training set. Six classification models (Stepwise Logistic Regression, LASSO, Random Forest, GBM, Neural Network, and CART) were trained and evaluated using ROC/AUC and other diagnostic metrics. DeLong's test was applied for pairwise AUC comparisons.

MRI-derived features, particularly from wavelet-transformed gldm and first-order matrices, showed strong predictive power, with several achieving > 0.75 AUC individually. The gradient boosting machine (GBM) outperformed all other models with an AUC of 0.998 and an accuracy of 95.6%. Random forest and CART also demonstrated high discriminative performance (AUCs of 0.990 and 0.914, respectively). Nine features were consistently ranked among the top 20 predictors across all models, suggesting robust modality-independent imaging biomarkers. DeLong's test confirmed statistically significant AUC differences between GBM and logistic regression models (p < 0.05).

The results of this pilot study suggest that radiomic analysis combining CT and MRI modalities, particularly when processed through ensemble machine learning methods, holds the potential to accurately predict RAS mutational status in CRLM. While promising, these findings should be interpreted with caution, considering the study's limitations, including the small patient cohort and its design. These factors highlight the need for prospective validation in larger, multicenter cohorts to confirm the generalizability of the models. Nevertheless, these preliminary results support the use of multiparametric radiomics as a potential non-invasive tool for preoperative molecular stratification.