RADT-14. COMPARISON OF MACHINE LEARNING ALGORITHMS IN DISTINGUISHING RADIATION NECROSIS FROM PROGRESSION OF BRAIN METASTASES TREATED WITH STEREOTACTIC RADIOSURGERY (SRS)

Abstract

Abstract PURPOSE To test the effectiveness of machine learning algorithms in distinguishing radiation necrosis (RN) from tumor progression (TP) using MRI radiomic features. METHODS Brain metastases were treated with SRS to a median dose of 18Gy. Lesions that showed evidence of progression on follow-up MRI were sampled surgically, and diagnoses confirmed by histopathology. Cases from 2 institutions were combined and randomly assigned for training (70%) and testing (30%). T1 post-contrast (T1c) and T2 fluid attenuated inversion recovery (T2 FLAIR) MRI were used for radiomic feature extraction (50 features each). Three subsets of radiomic features were obtained and tested: Signature #1 included 10 previously published features that correlated with diagnosis on T test; signature #2 and #3 included 5 and 12 features obtained through recursive elimination using random forest (RF) and support vector machine (SVM), respectively. Supervised machine learning models were trained using RF, SVM (radial kernel) and regularized discriminant analysis (RDA) algorithms based on all three radiomics signatures. Receiver operator characteristics (ROC) were compared between signatures and algorithms. RESULTS A total of 135 individual lesions (37 RN and 98 TP) were included. Signature #3 demonstrated the highest area under the curve in the training set (average AUC=0.98, vs 0.95 and 0.92 for signature #1 and #2), as well as the testing set (average AUC=0.83, vs 0.74 and 0.79 for signature #1 and #2). RF and SVM demonstrated similar performance in both training (average AUC 0.99-1) and testing datasets (average AUC 0.79-0.80) among all three signatures. Both RF and SVM were superior to RDA in performance (average training AUC 0.83, testing AUC 0.77). The greatest sensitivity (83%) and specificity (100%) in the testing set were achieved using signature #3 and SVM. CONCLUSION RF and SVM are effective in distinguishing RN from TP in a multi-institution dataset using radiomic signatures.