Session Item

Surgical resection is the recommended treatment in patients with large or symptomatic brain metastases (BM). Adjuvant radiation to the resection cavity improves local control. Still, there remains a risk of local failure (LF). Therefore, we aimed to develop and externally validate a radiomics-based prediction tool to determine patients at high risk of LF.

Data was collected within the “A Multicenter Analysis of Stereotactic Radiotherapy to the Resection Cavity of Brain Metastases” (AURORA) retrospective trial. In total, we collected data from 453 patients, of whom 321 had sufficient MR imaging and clinical data. The patients were divided into a training cohort of 237 patients from two centers and a multicentric external test cohort of 84 patients from four centers.
All BMs and their surrounding FLAIR-hyperintense regions were segmented. We extracted 104 radiomic features from contrast-enhanced T1-weighted sequences (T1-CE) and FLAIR sequences each. Furthermore, eight clinical features (primary diagnosis, control of primary, number of BMs, location, resection status, extracranial metastases, age, and Karnofsky index before RT) were included.
We compared five different feature sets: T1-CE, FLAIR, T1-CE+FLAIR, clinical and a T1-CE+FLAIR+clinical feature set. We trained three different learners on each feature set: random forest, extreme gradient boosting regression, and elastic net regression (ENR). The number of features was reduced by using Pearson intercorrelation and the Boruta algorithm on 50 bootstrap samples. Parameter tuning and model selection were performed using 20 iterations of five-fold cross-validation on our training set. The final models were trained on the whole training cohort with the best parameter set and tested on the external test set. We also tested the predictive value of BM volume (BMV) alone with a cox proportional hazard model (CPH). Performance was calculated using the concordance index (CI).

The ENR model ranked best in our internal cross-validation and was selected for external testing. The best performance in our multicentric external test cohort was achieved by ENR trained on the T1-CE+FLAIR+clinical feature set with a total of 15 chosen features by Boruta (2 clinical and 13 radiomic features) with a CI of 0.74. It outperformed the T1-CE and FLAIR-based models (CI: 0.70 and 0.63) as well as the clinical model (CI: 0.67). The T1-CE+FLAIR+clinical model stratified patients in Kaplan Mayer analysis on the test set significantly (p = 0.003). The predictor correlated significantly with BMV (r = 0.57, p < 0.01).  Adding BMV as a feature further improved the performance to 0.77. While BMV alone was highly predictive with a CI of 0.75 in the test cohort, it only achieved a CI of 0.53 in internal validation.

Our combined model predicted LF better than clinical features alone in an external multicenter test cohort. Patients with an increased risk of LF could benefit from intensified therapy regimens or follow-up frequencies.