A new radiomics-based approach for predicting brain tumor recurrence

The prediction and identification of tumor recurrence is important for planning follow-up and treatment strategies for patients with brain tumors. Radiomics has shown promise for recurrence prediction; however, its high-dimensional feature space increases the risk of overfitting and reduces generalizability, particularly in small-sample datasets. Conventional principal component analysis (PCA) can reduce dimensionality but complicates feature interpretation, while feature selection may be affected by multicollinearity among radiomic variables. In this study, Magnetic Resonance Imaging (MRI)-derived radiomic features were combined with clinical variables to predict brain tumor recurrence after radiotherapy. We evaluated an alternative radiomics workflow designed for limited-sample settings that identifies the most predictive feature family, incorporates a categorical feature based on primary tumor location, and applies rank-correlation-based hierarchical clustering within the selected feature family. A categorical scaling parameter was also introduced to reduce the dominance of one-hot encoded clinical variables over scaled numerical radiomic features. On the University of Mississippi Medical Center (UMMC) benchmark dataset, the proposed workflow achieved a higher held-out test Receiver Operating Characteristics Area Under Curve (AUC) of 0.91 than PCA (0.81) and mRMR (0.75) baselines and outperformed previously reported methods on the same dataset. These findings are preliminary and require external validation before clinical adoption.