Radiomics-based Machine Learning Approach to Predict Chemotherapy Responses in Colorectal Liver Metastases.

OBJECTIVES

This study explored the clinical utility of CT radiomics-driven machine learning as a predictive marker for chemotherapy response in colorectal liver metastasis (CRLM) patients.

METHODS

We included 150 CRLM patients who underwent first-line doublet chemotherapy, dividing them into a training cohort (n=112) and a test cohort (n=38). We manually delineated three-dimensional tumor volumes, selecting the largest liver metastasis for measurement, using pretreatment portal-phase CT images and extracted 107 radiomics features. Treatment response was classified as responder (complete or partial response) or non-responder (stable or progressive disease), based on the best overall response according to RECIST criteria, version 1.1. Employing Random Forest and Boruta algorithms, we identified significant features for responder-non-responder differentiation. Radiomics signatures were developed and validated in the training cohort using five-fold cross-validation, and performance was assessed using the area under the curve (AUC).

RESULTS

Among the patients, 91 (61%) were responders and 59 (39%) were non-responders. Variable selection with Boruta revealed three key parameters ("DependenceVariance," "ClusterShade," and "RunVariance"). In the training cohort, individual CT texture parameter AUCs ranged from 0.4 to 0.65, while the machine learning analysis incorporating all valid parameters exhibited a significantly higher AUC of 0.94 (p<0.01). The validation cohort also demonstrated strong predictive accuracy, with an AUC of 0.87 for treatment response.

CONCLUSIONS

This study highlights the potential of CT radiomics-driven machine learning in predicting chemotherapy responses among CRLM patients.