Feature-Based Machine Learning for Brain Metastasis Detection Using Clinical MRI

Brain metastases represent one of the most common intracranial malignancies, yet early and accurate detection remains challenging, particularly in clinical datasets with limited availability of healthy controls. In this study, we developed a feature-based machine learning framework to classify patients with and without brain metastases using multi-modal clinical MRI scans. A dataset of 50 subjects from the UCSF Brain Metastases collection was analyzed, including pre- and post-contrast T1-weighted images and corresponding segmentation masks. We designed advanced feature extraction strategies capturing intensity, enhancement patterns, texture gradients, and histogram-based metrics, resulting in 44 quantitative descriptors per subject. To address the severe class imbalance (46 metastasis vs. 4 non-metastasis cases), we applied minority oversampling and noise-based augmentation, combined with stratified cross-validation. Among multiple classifiers, Random Forest consistently achieved the highest performance with an average accuracy of 96.7% and an area under the ROC curve (AUC) of 0.99 across five folds. The proposed approach highlights the potential of handcrafted radiomic-like features coupled with machine learning to improve metastasis detection in heterogeneous clinical MRI cohorts. These findings underscore the importance of methodological strategies for handling imbalanced data and support the integration of feature-based models as complementary tools for brain metastasis screening and research.