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

Abstract

Background: Surgical resection is the standard of care for patients with large or symptomatic brain metastases (BMs). Despite improved local control after adjuvant stereotactic radiotherapy, the risk of local failure (LF) persists. Therefore, we aimed to develop and externally validate a pre-therapeutic radiomics-based prediction tool to identify patients at high LF risk.

Methods: Data were collected from A Multicenter Analysis of Stereotactic Radiotherapy to the Resection Cavity of BMs (AURORA) retrospective study (training cohort: 253 patients from 2 centers; external test cohort: 99 patients from 5 centers). Radiomic features were extracted from the contrast-enhancing BM (T1-CE MRI sequence) and the surrounding edema (T2-FLAIR sequence). Different combinations of radiomic and clinical features were compared. The final models were trained on the entire training cohort with the best parameter set previously determined by internal 5-fold cross-validation and tested on the external test set.

Results: The best performance in the external test was achieved by an elastic net regression model trained with a combination of radiomic and clinical features with a concordance index (CI) of 0.77, outperforming any clinical model (best CI: 0.70). The model effectively stratified patients by LF risk in a Kaplan-Meier analysis (P < .001) and demonstrated an incremental net clinical benefit. At 24 months, we found LF in 9% and 74% of the low and high-risk groups, respectively.

Conclusions: 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.

Keywords: artificial intelligence; brain metastases; local failure prediction; machine learning; radiomics.

Figure 1.

Summarized overview of our workflow. After manual and automatic definition of the volume of interest (VOI), we extracted 104 original features from each metastasis and edema segmentation. We reduced the number of features in each set with MRMR. Furthermore, we added up to 8 clinical features and combined all features into multiple different feature sets. The optimal number of features in each set was determined with a nested cross-validation. The optimal parameters for our selected learners were chosen based on a 5-fold cross-validation. The best parameters for each learner-feature-combination were tested in the external test cohort.

Figure 2.

Kaplan–Meier analysis. We created dichotomous predictions of the comb + pre-OP ENR model by using the 66th percentiles of the continuous risk ranks in the training cohort as cutoffs for patient stratification. There were 6 and 10 events in the low-risk group of 76 patients and the high-risk group of 23 patients, respectively. We found a significant difference in freedom from local failure (FFLF) between the predicted low- and high-risk groups (P < .001) in the multicenter external test cohort. After 24 months, we found a FFLF of 91% and 26% in the groups, respectively.

Figure 3.

Decision curve analysis (left) and calibration curve (right). Using the same groups as in Figure 2, we found a net benefit of our predictive model compared to treating all patients in the relevant threshold range from 5% to 30% through decision curve analysis (left). A decision model shows a clinical benefit if the respective curve shows larger net benefit values than reference strategies. The combination of radiomic features derived from T1-CE, FLAIR, and pre-OP features (comb + pre-OP) resulted in a higher net benefit compared to using only the clinical pre-OP features and treating all patients or none. The calibration curve on the right was created by transforming the continuous risk rank predicted by the best comb + pre-OP ENR model (in orange) and by the clinical pre-OP ENR model (in blue) to event probabilities at 24 months. Although both models seem to overestimate the actual risk of our patients, the comb + pre-OP model predicted the risk closer to the actual risk.