Radiomics for risk stratification in brain metastases: potential in melanoma brain metastases

Abstract

The identification of quantitative non-invasive imaging biomarkers, including radiomics, may complement molecular characterization and thereby improve clinical management of neuro-oncological patients. We aimed to identify imaging predictors with improved performance over clinical parameters to stratify patients with brain metastases into high and low-risk groups for overall survival (OS). 422 patients recruited by two neuro-oncological centers were included with first diagnosis of brain metastases from different primary tumors. From each patient, 15 clinical parameters and a total of 321 radiomic features extracted from cerebral MRI were employed in prediction models to classify patients into low- and high-risk groups for OS. The best performing model was a bootstrap aggregating model including only clinical features (test set: macro F-1 = 0.62, accuracy = 0.72), while the combined and radiomic datasets led to poorer results (test set: macro F-1 = 0.60, accuracy = 0.67; test set: macro F-1 = 0.62, accuracy = 0.52, respectively). However, in the subgroup of melanoma patients (n = 54), the radiomic dataset showed better predictive power over clinical and combined dataset (test set: macro F-1 score = 0.71, accuracy = 0.77). 80% and 67% of melanoma patients were correctly classified into the low- and the high-risk group for OS, respectively. Clinical features remained the strongest predictors of risk across patients with brain metastases from different primary tumors. Although, in melanoma patients, radiomic features provided better prediction of the survival outcome compared to clinical parameters alone.