Abstract
A combination of clinical and radiomic features predicted freedom from local failure (LF) better than any clinical feature set alone. Patients at high risk for LF may benefit from stricter follow-up routines or intensified therapy.
Introduction
Brain metastases (BMs) are a major complication of systemic cancers. While surgical resection followed by stereotactic radiotherapy (SRT) has improved patient outcomes, local failure—the return of the tumor at the original site—remains a persistent threat. Current clinical indices like the Graded Prognostic Assessment (GPA) are designed to estimate overall survival but lack the specificity to predict local control. Identifying high-risk patients before failure occurs would allow for intensified treatment (e.g., dose escalation) or stricter imaging schedules, while low-risk patients could be spared unnecessary scans.
Methods
Utilizing data from the multicenter AURORA study, researchers developed a machine learning model to predict local failure as a survival task. The workflow involved extracting 104 quantitative radiomic features—covering shape, intensity, and texture—from pre-therapeutic T1-weighted contrast-enhanced (T1-CE) and T2-FLAIR MRI sequences.
To ensure robustness, the team used a "minimum redundancy—maximum relevance" (MRMR) framework for feature selection and an Elastic Net Regression (ENR) model to account for censored data. The model was trained on 253 patients from two centers and then tested on a completely independent external cohort of 99 patients from five different centers across Germany and Switzerland.
Results
The combined Radiomics + Clinical model achieved a Concordance Index (CI) of 0.77 on the external test set, significantly outperforming the best clinical-only model (CI 0.70).
As shown in the Kaplan-Meier analysis, the model effectively stratified patients into risk groups. By the 24-month mark, the low-risk group had a local failure rate of only 9%, compared to 74% in the high-risk group. Decision curve analysis confirmed that this approach provides a net clinical benefit over standard strategies, particularly for clinicians prioritizing the early detection of recurrence.
Discussion
A key insight from this study is the importance of the tumor microenvironment; features extracted from the T2-FLAIR sequence (edema) added significant predictive value. Interestingly, the study found that tumor texture and shape (such as "flatness") were often more indicative of biological behavior than raw volume.
While manual segmentation remains the gold standard, the researchers also demonstrated the feasibility of automated workflows using neural networks, which achieved a respectable CI of 0.72. Limitations include the retrospective challenge of distinguishing local failure from radiation necrosis. Future work will focus on prospective trials to determine if risk-adapted therapy based on these models can improve overall patient survival. The authors have released their model as a web application to encourage further external validation.
Summary updated