BIOM-17. LEVERAGING ARTIFICIAL INTELLIGENCE TO DETECT SIGNATURES PREDICTING EARLY CEREBRAL RADIATION NECROSIS VIA TRACKING MULTIMODAL MRI RADIOMICS EVOLUTION: AI-SPECTRE

Abstract PURPOSE Radiation necrosis (RN) remains to be the primary dose-limiting toxicity of stereotactic radiosurgery (SRS), especially in combination with immunotherapy (IO). This debilitating adverse event is often confused with tumor regrowth, posing significant diagnostic and therapeutic challenges. This study employed magnetic resonance imaging (MRI) radiomics to quantify premanifest radiation-induced changes in irradiated brain metastases (BMs) and their value in predicting treatment response. METHODS AND MATERIALS Following IRB approval, we queried our single-institution dataset to identify adult patients with BMs managed with SRS and IO between 2006 and 2021. We collected patient, disease, and treatment variables. Specifically, per-BM outcomes were categorized as RN (radiographic and/or symptomatic), progressive disease (PD), or neither (NA). All patients had diagnostic pre- and post-SRS contrast-enhanced T1-weighted MRIs. Individual BMs were manually segmented in MIM Software (Beachwood, OH). 1061 hand-crafted radiomic features were computed per lesion using IBEX Software. Simple then multiple nominal logistic regressions (SLR and MLR) with Bonferroni correction ensued to model the risk of RN, PD, or NA using JMP software. RESULTS 92 patients with 301 BM of NSCLC, melanoma, and renal cell carcinoma primaries were included. All patients received brain SRS and IO. RN, PD, or NA occurred in 74 (24.6%), 75 (24.9%), and 152 (50.5%) BM, respectively. 39 radiomic features significantly changed post-SRT (p<0.01). Using MLR modeling, we created a 3-feature radiomic RN predictor (ROC AUC 0.71). The model included post-SRT BM surface area, texture (GLCM 3D Homogeneity), and pre-RT BM roundness. A 5-feature radiomic signature could predict RN and PD risk (ROC AUC 0.71) based on pre-SRT GLCM 3D texture feature, 3 post-SRT shape features (surface area, mean breadth, and roundness), in addition to delta-volume. CONCLUSION Pre-SRS, post-SRS, and percent changes of these radiomic features could be correlated to and potentially predict treatment outcomes. External validation and testing are required.