NIMG-46. MULTI-PARAMETRIC MRI RADIOMIC ANALYSIS USING DEEP LEARNING PREDICTS PERITUMORAL GLIOBLASTOMA INFILTRATION AND SUBSEQUENT RECURRENCE

Abstract

Abstract PURPOSE Glioblastoma (GBM) is the most aggressive and infiltrative brain tumor with a very poor prognosis which hasn’t significantly improved in over 20 years. The almost 100% recurrence rate is due to cancer infiltration beyond the tumor margins currently being targeted by standard of care. Prior studies used traditional supervised machine learning have shown great promise in predicting tumor infiltration beyond these margins. We hypothesize that deep learning methods can further improve such predictive maps and guide intensive, yet targeted and personalized therapies. METHODS MRIs from a cohort of 109 de novo GBM patients were collected from Hospital of the University of Pennsylvania. All the patients incorporated in this study had pre-operative multi-parametric MRI scans including T1, T1Gd, T2, T2-FLAIR, ADC, and underwent surgical resection followed by standard chemoradiation therapy and had pathologically confirmed recurrence. A novel, automated deep learning method, informed by results of prior studies using support vector machines, was constructed to train a patch-based ensemble CNN to identify regions of peri-tumoral cancer infiltration. Leave-one-out was used to evaluate the predictive value of this method, against pathology-proven subsequent recurrence. RESULTS Probability maps, representing the likelihood of tumor infiltration and eventual recurrence, were binarized using threshold of 50% cutoff, compared with actual recurrence on post-recurrence MRI scans. The average cross-validated accuracy was 92%, specificity was 93%, patient-based sensitivity was 78%, and odds-ratio among all patients was 12.95 (Estimated “hot spots” were 12.95 times more likely to present recurrence in the future). CONCLUSIONS This study demonstrates that Multi-parametric MRI pattern analysis using CNN based network can successfully predict tumor infiltration in peritumoral region for Glioblastoma patients. These suggest that new intensive, yet precisely targeted treatments can be developed, guided by AI-driven predictive maps of infiltration.