RADIOMICS TO PREDICT THE RESPONSE OF BRAIN METASTASES TO STEREOTACTIC RADIOTHERAPY

Abstract

Abstract AIMS Radiomics converts routinely obtained medical imaging into mineable, high-dimensionality data. We assessed its utility for incorporation into a multivariate model to predict the likelihood of response of brain metastases (BMs) to stereotactic radiotherapy (SRT). METHOD Patients treated with SRT over a ten-year period for BMs at three tertiary centres were retrospectively analyzed. SRT planning MRI scans were pre-processed, and radiomic features were extracted from each BM using PyRadiomics software. Follow up scans were analyzed for response and quantified by a modification of the Response Assessment in Neuro-Oncology Brain Metastases criteria. A radiomic signature predictive of BM response was created by a stepwise process involving correlation coeffcients and a binary least absolute shrinkage and se- lection operator regression analysis. A clinical predictive multivariate model for predicting response was com- pared against a combination clinical/radiomic model by McFadden’s pseudo-R-squared test. RESULTS 520 patients were included contributing a total of 1891 BMs. 1131 radiomic features were extracted from each BM. 12 features were used to construct a radiomic signature. Mean dose and BM volume were independent predictors of response in both the clinical and radiomics models. The radiomic signature was an independent prognostic predictor for response in the radiomic model (HR=1.095-1.231). Goodness of fit for predicting response was greater in the radiomics model (McFadden’s pseudo-R-squared score 0.243 vs 0.211), and statistically significant by likelihood ratio test (p=0.021). CONCLUSIONS This multi-centre retrospective analysis created a unique radiomic signature for predicting BM response to SRT and is the first study to incorporate clinical factors. A prospective validation study is planned.