P13.13.B INTERPRETABLE WHOLE SLIDE IMAGE PROGNOSTIC STRATIFICATION OF GLIOBLASTOMA PATIENTS FURTHERING DISEASE UNDERSTANDING

Abstract

Abstract BACKGROUND Glioblastoma is the most common malignant adult brain tumor, with a grim prognosis and heterogeneous morphology. Stagnant patient prospects over the last 20 years reflect our limited disease understanding. Robust prognostic patient stratification from whole slide images (WSI) using interpretable computational methods could improve disease understanding and patient management. MATERIAL AND METHODS Assessment of the TCGA-GBM and TCGA-LGG data collections by the 2021 WHO classification criteria of CNS tumors identified 188 glioblastoma (IDH-wt, grade 4) cases, divided into short (<9 months, n=94) and long (>13 months, n=94) survivors. An H&E-stained WSI from each patient is selected and comprehensively pre-processed to remove artifactual and non-tissue regions by partitioning them into 256x256 patches. An interpretability mechanism via a weakly supervised attention-based multiple-instance-learning algorithm is leveraged for prognostic stratification. The attention scores are visualized as a heatmap to identify the essential ROIs driving decisions. The model is evaluated on 10-fold monte carlo cross-validation, using training (80%), validation (10%), and test (10%) sets. RESULTS Quantitative evaluation revealed ValidationACCURACY=70%, TestACCURACY =62%, ValidationAUC= 0.75, and TestAUC=0.68. Heatmaps indicate weight given throughout histologically malignant areas for correctly predicted long survivors, including necrosis, hypercellularity, atypia, infiltration, and proliferation. Correct prediction of short survival weighted infiltrative features, including leptomeningeal involvement. For misclassified long-survivors, necrotic areas are not heavily weighted, and gemistocytic cells are seen in a densely packed rather than infiltrating arrangement. For some misclassified short-survivors, the hot regions are hypercellular; however, these are a minority of the specimen, which predominantly showed a less cellular, fibroblastic, spindled appearance. CONCLUSION The interpretation of algorithmic decisions through qualitative visualization of heatmaps can offer insights into histologic regions with prognostic relevance. In the case of glioblastoma, the histologic appearance of long survivors based on a neuropathologist's observation suggests aggressive tumors traditionally considered indicative of poor prognosis. The algorithm's ability to accurately predict long survival based on these areas indicates that further investigation of these features may be warranted for improved prognostication through histologic assessment. Our findings support the use of data-driven interpretation of algorithmic decisions to identify morphological patterns with prognostic significance, which would further our understanding of glioblastoma.