NIMG-19. SYNTHETIC ISOCITRATE DEHYDROGENASE-MUTANT GLIOBLASTOMAS FROM GENERATIVE ADVERSARIAL NETWORK PROVIDE MORPHOLOGIC VARIABILITY AND DIAGNOSTIC PERFORMANCE SIMILAR TO REAL DATA: DEVELOPMENT AND VALIDATION

Abstract

Abstract BACKGROUND Generative adversarial network (GAN) creates synthetic MRI data that may provide morphologic variability to assess molecular characteristics of glioblastomas. PURPOSE To investigate the ability of GAN-based generation of isocitrate dehydrogenase (IDH)-mutant glioblastomas to provide morphologic variability and improve molecular prediction. METHODS GAN was retrospectively trained on 110 IDH-mutant high-grade gliomas. Paired contrast-enhanced T1-weighted and FLAIR synthetic MRI data were generated. Diagnostic models were developed from 80 IDH-wild type glioblastomas and 38 IDH-mutant patients, (real model), 38 IDH-mutant GAN-generated synthetic data (synthetic model), or both combined (augmented model). Two neuroradiologists independently assessed real and morphologic characteristics of contrast-enhancement patterns, the presence of necrosis, and margins and type of non-enhancing region. Significant predictors of IDH mutation were selected from multivariable logistic regression, and diagnostic performance was validated in 44 separate patients, 33 with IDH-wild type and 11 with IDH-mutant glioblastomas. RESULTS Synthetic IDH-mutant glioblastomas were similar to real tumors on Turing tests, with an area under the curve (AUC) of 0.67–0.71. Significant predictors of a more frontal or insular location (odds ratio [OR], 1.34 vs. 1.52; highest P = .04) and distinct non-enhancing tumor margins (OR, 2.68 vs. 3.88; P < .001) were similar for the real and synthetic models, with the synthetic (AUC, 0.958) and augmented (AUC, 0.899) models having higher diagnostic performance than the real model (AUC, 0.864) in the training set. The diagnostic accuracy was higher for the synthetic and augmented models (90.9% [40/44] each for both observers) than for the real model (84.1% [37/44] for one observer and 86.4% [38/44] for the other). CONCLUSIONS The GAN-based synthetic images yield morphologically variable IDH-mutant glioblastomas and may be useful as realistic training sets.