NIMG-22. COMPARISON OF NON-INVASIVE, ML-GENERATED WHOLE-BRAIN SPATIAL RADIOPATHOMIC MAPS DERIVED FROM ANATOMICAL AND DIFFUSION-WEIGHTED MRI ON PATIENTS WITH NEWLY DIAGNOSED-GLIOMA

Abstract

Abstract INTRODUCTION Spatial heterogeneity in the glioma microenvironment is difficult to capture through singular biopsy samples. Identification of malignant tumor regions can help guide diagnosis and treatment planning. This study compares radiopathomic maps from 2 strategies derived from anatomical and diffusion-weighted MRI: 1) a support vector machine model trained on an MRI dataset with tissue samples of known spatial coordinates to generate spatial maps of probability of KI-67 and cellularity, and 2) a bagged random-forest model derived from 40+ person glioma brain bank of autopsy tissue coregistered to end-of-life MRI to generate spatial maps of cellularity and tumor probability (TPM). We hypothesize that TPMs from autopsy samples will correlate with KI-67 predictions, while cellularity values among methods will correlate. METHODS Spatial maps of tumor pathology from autopsy- and tissue sample-based models were generated on 55 patients newly-diagnosed with glioblastoma that were scanned with anatomical, diffusion-weighted, and metabolic imaging. Mean values were extracted from each map within the T2-hyperintense, non-enhancing lesion (NEL), the contrast-enhancing lesion (CEL), normal-appearing white matter (NAWM), and 5mm diameter spheres centered at the location where tissues were taken. Spearman correlation analysis was used to elucidate the strength of the linear relationship within the selected regions between the two maps. RESULTS Correlations between KI-67 and TPM values were highest and most significant at the tissue sample locations and CEL (ρ=0.306, p=0.002 and ρ=0.312, p=0.044 respectively). Cellularity maps generated from the autopsy-based model showed no significant correlation with newly diagnosed cellularity values at tissue sample locations, potentially highlighting the influence of treatment. CONCLUSION This study examines the current state of ML-generated spatial mapping, highlighting similarities within CEL, but need to account for the effects of subsequent treatment and progression when comparing tumor beyond the CEL between models derived between different time points.