TMOD-34. A RADIOMIC SIGNATURE OF INFILTRATION IN PERITUMORAL EDEMA PREDICTS SUBSEQUENT RECURRENCE IN GLIOBLASTOMA

Abstract

Abstract BACKGROUND The malignant parenchyma in glioblastoma extends beyond the enhancing borders of tumor on postcontrast T1-weighted magnetic resonance imaging (MRI), which is the primary target of treatment. Such non-enhancing tumor invasion into the peri-tumoral edema (PED) is, however, not usually distinguishable on conventional MRI. The aim of this study was to evaluate pre-operative MRI in the PED to assess whether areas of tumor infiltration and early recurrence can be detected. METHODS A cohort of 90 de novo glioblastoma patients from a single institution (Penn) was selected. All included patients had preoperative multi-parametric MRI (mpMRI;T1,T1Gd,T2,T2-FLAIR,ADC), underwent initial gross-total-resection followed by standard chemoradiation, and had pathologically-confirmed recurrence. An extensive panel of handcrafted features, including shape, volume, intensity distributions, texture, was extracted from mpMRI scans. Predictive modeling for estimation of PED infiltration was performed using sequential feature selection approach designed with a support vector machine classifier and through a Leave-one-out cross-validation approach in the Penn cohort. Generalizability of the model was evaluated by applying it on a cohort of 20 patients from a second institution (Case), and predicted probability distributions in PED were compared in both the cohorts. RESULTS Spatial probability maps, representing the likelihood of tumor infiltration and eventual recurrence, were binarized at 50% cutoff, and compared with actual recurrence on post-recurrence scans. The cross-validated accuracy of our model within Penn cohort was 81.35% (odds-ratio=3.62, sensitivity/specificity=78.26/81.35). The model trained on the Penn cohort, when applied on the Case cohort, produced almost similar intensity distribution in the PED, suggesting that the method has potential for robust performance across institutions. The comparison of intensity distributions revealed higher ADC and T2-FLAIR in non-recurrent regions compared to recurrent ones. CONCLUSION Multi-parametric pattern analysis of mpMRI across multiple institutions generates similar, accurate estimates of spatial extent and patterns of recurrence in PED, which may guide strategies for treatment intensification in glioblastoma.