NIMG-25. IMPROVING THE NONINVASIVE CLASSIFICATION OF GLIOMA GENETIC SUBTYPE WITH DEEP LEARNING AND DIFFUSION-WEIGHTED IMAGING. (12th November 2021)

• Record Type: Journal Article
• Authors: Cluceru, Julia; Phillips, Joanna; Molinaro, Annette; Interian, Yannet; Luks, Tracy; Alcaide-Leon, Paula; Nair, Devika; LaFontaine, Marisa; Shai, Anny; Chunduru, Pranathi; Pedoia, Valentina; Villanueva-Meyer, Javier; Chang, Susan; Lupo, Janine
• Is Part Of: Neuro-oncology. Volume 23: Supplement 6(2021)
• Page Start: vi133
• Page End: vi134
• Publication Date: 2021-11-12
• Subjects: Brain Neoplasms -- Periodicals; Brain -- Tumors -- Periodicals; Brain -- Cancer -- Periodicals; Nervous system -- Cancer -- Periodicals; 616.99481
• DOI: 10.1093/neuonc/noab196.525
• Languages: English
• ISSNs: 1522-8517

Abstract

Abstract: In contrast to the WHO 2016 guidelines that use genetic alterations to further stratify patients within a designated grade, new recommendations suggest that IDH mutation status, followed by 1p19q-codeletion, should be used before grade when differentiating gliomas. Although most gliomas will be resected and their tissue evaluated with genetic profiling, non-invasive characterization of genetic subgroup can benefit patients where surgery is not otherwise advised or a fast turn-around is required for clinical trial eligibility. Prior studies have demonstrated the utility of using anatomical images and deep learning to distinguish either IDH-mutant from IDH-wildtype tumors or 1p19q-codeleted from non-codeleted lesions separately, but not combined or using the most recent recommendations for stratification. The goal of this study was to evaluate the effects of training strategy and incorporation of Apparent Diffusion Coefficient (ADC) maps from diffusion-weighted imaging on predicting new genetic subgroups with deep learning. Using 414 patients with newly-diagnosed glioma (split 285/50/49 training/validation/test) and optimized training hyperparameters, we found that a 3-class approach with T1-post-contrast, T2-FLAIR, and ADC maps as inputs achieved the best performance for molecular subgroup classification, with overall accuracies of 86.0%[CI:0.839, 1.0], 80.0%[CI:0.720, 1.0], and 85.7%[CI:0.771, 1.0] on training, validation, and test sets, respectively, and final test class accuracies of 95.2%(IDH-wildtype), 88.9%(IDH-mutated, 1p19qintact), and 60%(IDHmutated, 1p19q-codeleted). Creating an RGB-color image from 3 MRI images and applying transfer learning with a residual network architecture pretrained on ImageNet resulted in an 8% averaged increase in overall accuracy. Although classifying both IDH and 1p19q mutations together was overall advantageous compared with a tiered structure that first classified IDH mutational status, the 2-tiered approach better generalized to an independent multi-site dataset when only anatomical images were used. Including biologically relevant ADC images improved model generalization to our test set regardless of modeling approach, highlighting the utility of incorporating diffusion-weighted imaging in future multi-site analyses of molecular subgroup.