NIMG-23. AI-BASED CONNECTED COMPONENT MARKERS OF BRAIN PERITUMORAL MICROENVIRONMENT USING WATER RESTRICTION INFORMATION

Abstract

Abstract PURPOSE Glioblastoma is the most aggressive adult brain tumor, with heterogeneous neoplastic cell peritumoral infiltration. Characterization of this infiltrative peritumoral heterogeneity is an unmet clinical need that could contribute to strengthening our understanding of this disease. We propose novel AI-based markers of infiltration using deep-learning (DL) based on water restriction caused by infiltration, identified by diffusion tensor imaging (DTI). These markers could contribute to precision/personalized medicine, towards influencing clinical decision-making, including planning for biopsies, surgery, and radiation. METHOD: We automatically extracted peritumoral patches from free water volume fraction maps (FW-VF) of a retrospective cohort of 44 brain metastases and 66 glioblastomata patients and labelled them as high- and low- free water, respectively. An AI/DL model was then trained on these patches to distinguish differences in water restriction. Our trained AI/DL model was then applied on FW-VF of 264 hold-out glioblastoma patients (survival:0.43-76.9 months, age:21-88, 104 females) to generate a peritumoral microenvironment index (PMI) map quantifying infiltrative heterogeneity. Connected components (CCs) of high PMI values were calculated and their descriptive characteristics of size, number, shape, directionality, and spatial location, were extracted as AI-based markers. Gaussian mixture model clustering was then applied on these markers to determine if their representative infiltrative peritumoral heterogeneity can capture overall survival differences, by partitioning the patients into three clusters: low, moderate, and high risk. RESULTS The log-rank test yielded significant differences (p< 10-5) between low- and high-risk patients, (HR= 0.47, 95% CI:0.34-0.65; P< 0.005). Average PMI values were significantly greater in high-risk patients (P< 0.05). CONCLUSION We introduced novel AI-based markers of infiltration in the peritumoral microenvironment, using information of water restriction extracted from DTI. Our proposed markers can capture overall survival differences, based on the patterns of infiltration using DTI-based characterization of the water restriction, that show promise as clinically relevant prognostic biomarkers.