Radiomics-based prediction of local control in patients with brain metastases following postoperative stereotactic radiotherapy

Josef A. Buchner, Florian Kofler, Michael Mayinger, Sebastian M. Christ, Thomas Brunner, Andrea Wittig, Bjoern Menze, Claus Zimmer, Bernhard Meyer, Matthias Gückenberger, Nicolaus Andratschke, Rami A. El Shafie, Jürgen Debus, Susanne Rogers, Oliver Riesterer, Katrin Schulze, Horst Jürgen Feldmann, Oliver Blanck, Constantinos Zamboglou, Konstantinos Ferentinos, Angelika Bilger-Zähringer, Anca-Ligia Grosu, Robert Wolff, Marie Piraud, Kerstin A Eitz, Stephanie E. Combs, Denise Bernhardt, Daniel Rueckert, Benedikt Wiestler, Jan C. PeekenNeuro-Oncology 2024 · DOI

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.

Overview of the radiomics workflow, from VOI definition to external validation.

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).

Kaplan-Meier curves demonstrating effective stratification of patients into high- and low-risk groups for local failure.

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

Abstract

A combination of clinical and radiomic features predicted freedom from LF better than any clinical feature set alone. Patients at high risk for LF may benefit from stricter follow-up routines or intensified therapy.
BibTeX
@inproceedings{buchner2024radiomics,
  author = {Josef A. Buchner and Florian Kofler and Michael Mayinger and Sebastian M. Christ and Thomas Brunner and Andrea Wittig and Bjoern Menze and Claus Zimmer and Bernhard Meyer and Matthias Gückenberger and Nicolaus Andratschke and Rami A. El Shafie and Jürgen Debus and Susanne Rogers and Oliver Riesterer and Katrin Schulze and Horst Jürgen Feldmann and Oliver Blanck and Constantinos Zamboglou and Konstantinos Ferentinos and Angelika Bilger-Zähringer and Anca-Ligia Grosu and Robert Wolff and Marie Piraud and Kerstin A Eitz and Stephanie E. Combs and Denise Bernhardt and Daniel Rueckert and Benedikt Wiestler and Jan C. Peeken},
  title = {{Radiomics-based prediction of local control in patients with brain metastases following postoperative stereotactic radiotherapy}},
  booktitle = {Neuro-Oncology},
  year = {2024},
  doi = {10.1093/neuonc/noae098},
}
Figure