Abstract 2474: Deep learning-based tissue segmentation enables quantitative assessment and identifies dynamic markers of response from routine staining in glioblastoma

Abstract Glioblastoma (GBM), the most common and aggressive primary brain tumor in adults, has a highly heterogeneous histopathologic profile on routine H&E staining with challenging traditional morphology assessment. The reliability of morphological features in reflecting response to therapy is debated, as current assessment approaches remain inherently qualitative with a lack of dynamic quantifications after treatment. To address this, we aimed to (1) train an interpretable deep convolutional neural network classifier for semantic segmentation of tissue regions in GBM whole-slide images (WSIs) from routine H&E staining, (2) establish a quantitative baseline of morphological variations after standard-of-care treatment (radiotherapy/temozolomide) in a paired clinical cohort, and (3) assess this approach’s utility in identifying markers of response. Therefore, we annotated 3, 545 region-level ground truth labels in a supervised approach and trained a DenseNet classifier on the HALO AI platform to segment seven tissue classes in GBM: TC1 (Geographic necrosis), TC2 (Palisading necrosis), TC3 (Mesenchymal tissue), TC4 (Hemorrhage), and TC5 (Parenchyma), subdivided into TC5.1 (Heavy infiltration) and TC5.2 (Mild infiltration). The classifier achieved high performance on an independent set of 828 annotations (precision = 0.88, recall = 0.84, F1-score = 0.86), and digital quantifications strongly correlated with manual estimations extracted from clinical pathology reports (p < 0.001). We then identified a retrospective longitudinal cohort of 142 IDH-wildtype GBM patients and scanned the corresponding 912 whole-slide images of tissue from diagnostic resection paired with tissue from recurrence after standard-of-care treatment. Between the two time points, quantifications by the applied classifier showed significant increases in the classes of mild infiltration (+87%) and geographic necrosis (+75%), coupled with parallel decreases in heavy infiltration (-31%) and palisading necrosis (-45%) (p < 0.01). On recurrence, the ratio of heavy to mild infiltration (TC5.1:TC5.2) allowed significant stratification of overall survival, and rapid progressors had significantly larger areas of geographic necrosis. On multivariable analysis, dynamic increase in mild infiltration and decrease in mesenchymal tissue were independently associated with reduced risk of poor outcome. Taken together, our AI-based tissue segmentation classifier enables an automated quantitative assessment and allows interpretable clinical response stratification from routine staining, providing a potential enhancement of GBM diagnostic reporting. Citation Format: Georges Ayoub, Dina Elharouni, Connor Bossi, Seth Malinowski, Prem Prabhakar, Constantin Möller, Muhammad Shaban, Anurag Jayant Vaidya, Julia Japo, Andrew Zhang, Andres Santos, Drew F.K. Williamson, David Meredith, Faisal Mahmood, Keith L. Ligon. Deep learning-based tissue segmentation enables quantitative assessment and identifies dynamic markers of response from routine staining in glioblastoma [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 2474.