AI in neuro-oncology: predicting EGFR amplification in glioblastoma from Whole Slide Images using Weakly Supervised Deep Learning

Glioblastoma (GBM), the most aggressive type of brain cancer, is notorious for its resistance to treatments due to its rapid growth and high degree of heterogeneity. Epidermal Growth Factor Receptor (EGFR) is an important diagnostic, prognostic, and therapeutic biomarker in GBM. The current gold standard diagnosis of EGFR detection relies on immunohistochemistry (IHC), Fluorescence in Situ Hybridization (FISH), and sequencing analysis of surgical samples with a turnaround time of several days to weeks, not to mention the laboratory costs related to the analyses. To improve the efficiency and cost-effectiveness of EGFR detection, we propose an alternative solution based on Whole Slide Image (WSI) and artificial intelligence (AI) methods, specifically multiple instance learning (MIL) frameworks. In this study, we conducted a comprehensive analysis of two public datasets, TCGA and CPTAC, with three MIL frameworks, Clustering Constrained Attention Multiple Instance Learning (CLAM), Double–Tier Feature Distillation Multiple Instance Learning (DTFD), and Transformer-based Correlated MIL for WSI Classification (TransMIL), to predict the EGFR status in GBM. With evaluation using 5–fold cross-validation, the CLAM model achieved 0.736 ± 0.181 and 0.799 ± 0.164 area under the curve (AUC) on TCGA and CPTAC, respectively, outperforming the state-of-the-art model (AUC of 0.71 ± 0.031). This study demonstrates the effectiveness of MIL models in the classification of EGFR biomarkers with a high potential to shift the paradigm from the current gold standard to the more efficient and cheaper WSI and AI workflows.