Machine learning models using 18F-FDG PET/CT radiomics for RAS mutation prediction and prognostic stratification in colorectal cancer.

OBJECTIVE

To evaluate a machine learning (ML) model that integrates clinical data and 2-deoxy-2-[18F]fluoro-D-glucose (18F-FDG)-PET radiomic features for predicting RAS mutation status and prognosis in patients with colorectal cancer (CRC).

METHODS

This retrospective study included 90 patients (mean age, 64 years, 60 men) with CRC who underwent pretreatment 18F-FDG-PET/CT. Radiomic features were extracted from PET images. Ten clinical variables and 49 radiomic features were analyzed. Using the AutoGluon ML framework, 3 models (clinical, radiomics, and combined) were developed with 10-fold cross-validation. RAS mutation prediction was evaluated by the AUC, and interpretability was assessed using SHAP analysis. Survival outcomes were evaluated using a RSF model, and risk scores for disease progression were calculated based on true RAS and ML-predicted RAS models.

RESULTS

The radiomics ML model demonstrated the highest performance with an AUC of 0.675. SHAP analysis identified NGTDM_Complexity, kurtosis, and skewness as key contributors. For survival analysis, the C-indices based on true and ML-predicted RAS status were 0.685 and 0.675, respectively. A strong correlation was observed between the risk scores derived from the 2 models (r = 0.99, ρ = 0.98). Kaplan-Meier analysis demonstrated clear separation between high- and low-risk groups for both models (log-rank P < .01), with comparable survival curves.

CONCLUSION

Although the radiomics-based ML model demonstrated moderate performance in predicting RAS mutation status, it provided prognostic stratification comparable to that based on true genetic profiles.

ADVANCES IN KNOWLEDGE: 18F-FDG PET-based radiomics with ML may have potential as a noninvasive approach for genetic profiling and prognostic assessment in CRC, although further validation is required.