NEIM-06 COMBINING CLINICAL VARIABLES AND RADIOMIC FEATURES TO HELP DISTINGUISH RADIATION NECROSIS FROM TUMOR IN PATIENTS WITH MELANOMA BRAIN METASTASES TREATED WITH RADIOSURGERY

Abstract

Abstract BACKGROUND Following Gamma Knife SRS (GK-SRS), the conventional imaging characteristics of radiation necrosis (RN) mimic those of tumor progression, introducing considerable uncertainty in diagnosis. Previous studies have identified clinical variables associated with RN; however, diagnosis primarily relied on interpretation of imaging with only a minority confirmed using the gold standard of pathological examination. Furthermore, the cohorts of these studies included a mix of primary histologies. PURPOSE To identify the combination of clinical variables and radiomic features most predictive of RN in patients with melanoma brain metastasis (BM) with GK-SRS in order to train a machine learning classifier to distinguish RN from tumor progression. METHODS We retrospectively studied 86 patients with a melanoma BM that received initial GK-SRS followed by resection, thereby pathologically confirming tumor or RN. Clinical variables including lesion volume, age at surgery, GK-SRS dose, lesion hemorrhage, lesion location, gender, BM velocity, and drug therapy type were obtained from chart review. We extracted radiomic features from contrast-enhanced T1-weighted MR images using PyRadiomics. A consensus clustering algorithm identified representative radiomic features. Non-parametric hypothesis testing was performed on the clinical variables and representative radiomic features. RESULTS Of the 86 patients, 17 (19.8%) patients exhibited RN and 69 exhibited tumor progression. Lesion volume was associated with development of RN (p = 0.038<0.05) with a median volume of 1.5 cc (0.01-26.71 cc). Clustering analysis identified seven representative radiomic features; five were found to have statistically significant association with development of RN. CONCLUSION In this dataset with pathologically confirmed diagnoses in a histologically homogeneous patient cohort, we reproduced previously reported findings that the clinical variable of lesion volume is associated with RN and we identified several radiomic features associated with RN in patients with melanoma BM. We are using these variables and features to train a machine learning classifier to distinguish RN from tumor.