Deep learning-based tumor segmentation and radiogenomic model for predicting EGFR amplification and assessing intratumoural heterogeneity in glioblastoma

Jianpeng Liu, Chuyun Shen, Shufan Jiang, Yanfei Wu, Jiaqi Tu, Yifang Bao, Haiqing Li, Na Wang, Ying Liu, Ji Xiong, Xueling Liu, Yuxin Li

Abstract: To predict epidermal growth factor receptor (EGFR) amplification and explore the intratumoural heterogeneity of glioblastoma (GBM) using deep learning segmentation-based radiogenomics. A total of 654 patients were included from multiple datasets, divided into a training cohort, an internal validation cohort, and an external validation cohort. Tumor regions were segmented into contrast-enhancing tumour, necrotic non-enhancing core, peritumoural oedema based MR imaging, and radiomics features were extracted. Segmentation accuracy was quantified by the Dice similarity coefficient (DSC) with respect to manual segmentation. Five classifiers including Random Forest (RF), Support Vector Machine (SVM), Logistic Regression (LR), Neural Network (NN) and Decision Tree (DT) were evaluated for EGFR amplification prediction. Next, gene expression data from The Cancer Genome Atlas were analysed to identify differentially expressed genes related to EGFR amplification. Gene set enrichment analysis were employed to identify gene modules and enriched biological pathways associated with EGFR amplification. The segmentation performance was validated on two independent validation cohorts, achieving a mean DSC of 0.952 ± 0.026 and 0.961 ± 0.034, respectively.1409 radiomics features were respectively extracted from the the contrast-enhanced T1-weighted imaging images, thirty-seven signatures were identified through feature selection, leading to the development of a robust classification model. The Random Forest model demonstrated superior performance in predicting EGFR amplification in glioblastoma, with AUCs of 0.946 (training cohort), 0.917 (internal validation cohort), and 0.851 (external validation cohort). Radiomics signatures showed significant correlations with hub genes (|R| = 0.27–0.53, P < 0.05). The radiomics model showed robust performance in predicting EGFR amplification in glioblastoma, providing a non-invasive approach to support molecular diagnosis. The findings suggest that EGFR amplification is associated with distinct molecular characteristics, underscoring the heterogeneity of GBM and offering potential insights for individualized treatment planning.