EPCO-15. LINKING HISTOLOGICAL GLIOBLASTOMA PHENOTYPES TO TRANSCRIPTIONAL SUBTYPES AND PROGNOSIS USING DEEP LEARNING. (14th November 2022)

• Record Type: Journal Article
• Authors: Roetzer-Pejrimovsky, Thomas; Kiesel, Barbara; Nenning, Karl-Heinz; Klughammer, Johanna; Rajchl, Martin; Bock, Christoph; Hainfellner, Johannes; Baumann, Bernhard; Langs, Georg; Woehrer, Adelheid
• Is Part Of: Neuro-oncology. Volume 24(2022)Supplement 7
• Page Start: vii118
• Page End: vii119
• Publication Date: 2022-11-14
• Subjects: Brain Neoplasms -- Periodicals; Brain -- Tumors -- Periodicals; Brain -- Cancer -- Periodicals; Nervous system -- Cancer -- Periodicals; 616.99481
• DOI: 10.1093/neuonc/noac209.450
• Languages: English
• ISSNs: 1522-8517

Abstract

Abstract: Deep learning is increasingly utilized in medical image analysis and shows promise to substantially inform treatment decisions in clinical practice. While it finds broad applications in radiology or genomics, there are still relatively few applications of deep learning to the analysis of histological images. A disease in which a better understanding of underlying phenotype-genotype correlations seems crucial is glioblastoma, the most common malignant brain tumor in adults that features a dismal prognosis and limited therapeutic options. Using a large cohort of 276 glioblastoma patients with available histological, immunohistochemical, clinical and molecular data, we trained an Xception convolutional neural network to predict patient survival and transcriptional subtypes on histological images alone. We used 5-fold cross validation to evaluate the performance of the trained network. We achieve a high accuracy for predicting overall survival on univariable (p < 0.001) and multivariable (p = 0.01) analysis. Moreover, we observe a high accuracy for predicting the predominant transcriptional subtype with an AUC of 0.84, 0.81, and 0.71 for classical, mesenchymal and proneural subtypes, respectively (all p < 0.001). Finally, we plot heatmaps of the predicted transcriptional subtype and risk scores to directly link them to the histological phenotypes and the tumor microenvironment. Reassuringly, we find an association of the mesenchymal subtype with the perinecrotic niche and of the proneural subtype with the infiltrating edge. In conclusion, we introduce an easily adaptable convolutional neural network model for working with histological data, which can accurately predict outcomes and transcriptional subtypes in glioblastoma.