Prediction of Brain Metastasis Response to Stereotactic Radiosurgery Using MRI and Machine Learning: Effects of Primary Cancer Site and Metastasis Volume

Abstract

<h3>Purpose/Objective(s)</h3> Pre-treatment prediction of brain metastasis (BM) response to stereotactic radiosurgery (SRS) would provide valuable insights to clinicians during treatment selection and planning. Studies combining magnetic resonance imaging (MRI), radiomics and machine learning (ML) have shown the prediction of SRS outcomes. BM volume and primary cancer site are known to influence SRS outcomes, but no radiomics ML studies have analyzed these effects. We hypothesize that primary cancer site and metastasis volume effects will be present in the prediction performance of radiomics ML models. <h3>Materials/Methods</h3> A retrospective study recruited 99 BM patients (n = 123 BMs) treated with linear accelerator SRS. The study endpoint was size-based progression after SRS, with pseudo-progression effects accounted for. 107 radiomic features were extracted from the pre-treatment T1w contrast-enhanced MRI for each BM gross tumor volume planning structure. 12 clinical features were collected encompassing demographics, cancer status and SRS prescription. A random decision forest ML model used the radiomic and clinical features as input to produce a predicted SRS response. 250 bootstrapped dataset samples were used to reduce variability. <h3>Results</h3> SRS response of BMs with a melanoma or renal primary were found to be predicted most accurately with an area under the receiver operating characteristic curve (AUC) of 0.76 and 0.69, respectively. Colorectal and lung primary BMs achieved lower AUC values of 0.62 and 0.61, while breast primary BMs achieved AUC 0.48. The upper half of the presented table reports AUC, false negative rate (FNR), and false positive rate (FPR) for the model performance on all BMs, but also for only BMs <7.5 cc versus BMs >7.5 cc. SRS outcomes for BMs <7.5 cc were predicted more reliably compared to BMs >7.5 cc, leading to an ineffective model for clinical use. To counter this effect, features highly correlated to BM volume were removed and the experiment repeated. As shown in the table's lower half, overall performance remained largely unchanged while the volume effect was removed. <h3>Conclusion</h3> We have demonstrated that primary cancer site and BM volume are both relevant factors when assessing a ML model. We first conclude that ML model performance is dependent on the underlying distribution of the primary cancer populations. We also conclude that the correlation of radiomic features with BM volume creates a ML model that is inaccurate for BMs >7.5 cc. It is therefore imperative for future studies to control for these effects. Encouragingly, we conclude that more balanced model performance can be achieved by removing volume-correlated radiomic features.