NIMG-76. INCORPORATING IMAGING FEATURES OF NORMAL BRAIN TO ENHANCE SPATIALLY DIFFERENTIATING TREATMENT-INDUCED EFFECTS FROM RECURRENT GLIOMA WITH AI

Abstract INTRODUCTION Although AI-based approaches have been applied to discriminate between tumor recurrence (rTumor) and treatment-induced effects (TxE) from chemoradiotherapy in patients with recurrent glioma, these models typically either do not account for within-lesion mixtures of rTumor and TxE or generalize to independent test-sets with >80% accuracy. We hypothesize that incorporating imaging features of normal-appearing brain (NAB) into existing AI-based models that leverage tissue-samples with known locations on MRI and physiological imaging will significantly improve performance for distinguishing regions of TxE from rTumor and allow for visualization of rTumor beyond the T2-lesion. METHODS Using 254 pathology-confirmed tissue-samples with coordinates on diffusion-weighted, perfusion-weighted, and anatomical MRI from 135 glioma suspected of progression, we modified previously-developed binary, AI-based prediction models to learn multi-class targets consisting of rTumor, TxE, and contralateral NAB-regions (3/patient) with imaging signatures of CSF, white-matter, and grey-matter (659 total 10mm-isotropic samples). The multi-class ensemble model was trained with 4 unique repeats of constrained-5-fold-cross-validation and evaluated on a held-out test set (30 patients). This process was repeated 20 times, and the performance was compared to corresponding binary AI-based models. RESULTS NAB was easiest to distinguish from individual and combined classes (AUC-ROCs>0.98). Multi-class model performance for predicting TxE increased by 27% to 0.93±0.03 (p<0.00001) when discriminating between NAB and rTumor as compared to just rTumor with the binary models. While TxE vs rTumor alone was most difficult to predict with the multi-class model (AUC-ROC=0.73) achieving equivalent performance to binary models, there was 67% less variance in test performance. This approach allowed for the quantification of whole-brain probability maps of rTumor, TxE, and NAB along with a combined 3-channel colormap, facilitating the visualization of rTumor predictions beyond the T2-lesion. CONCLUSIONS Including NAB imaging features improved performance of AI-based models of rTumor/TxE, while expanding the predicted area to produce more interpretable whole-brain probability maps. Current work is evaluating their utility in prospectively guiding sampling of rTumor tissue and predicting survival.