Abstract

Glioblastoma is the most prevalent and aggressive primary brain malignancy, characterized by rapid progression and pronounced micro- and macroscopic heterogeneity. This biological variability is shaped by multiple factors, including tumor cell density, infiltration of adjacent healthy tissue, and distinct gene expression signatures. Among these molecular drivers, mutations in the EGFR gene are closely linked to shorter recurrence intervals and diminished overall survival in patients with glioblastoma. Non-invasive imaging modalities such as MRI offer substantial potential for inferring EGFR mutation status, thereby minimizing the risks associated with surgical biopsies and sampling variability.
In this study, we propose an artificial intelligence–based decision support system (DSS) designed to classify EGFR mutations in glioblastoma patients through automated segmentation of tumorous regions on MRI. The system integrates deep neural network architectures—Inception ResNet‑v2, DenseNet‑121, and ResNet‑50—trained on a dedicated glioblastoma dataset obtained from Memorial Hospital in Istanbul. Three MRI input configurations were evaluated: expert-segmented tumor regions, unsegmented scans, and scans without visible tumor presence. Model selection was optimized using information criteria (IC), ensuring a principled balance between predictive accuracy and architectural complexity.
Among the evaluated models, DenseNet‑121 demonstrated the highest performance, achieving accuracy scores of 0.952, 0.942, and 0.938 for expert-segmented, unsegmented, and tumor-absent inputs, respectively. Precision and recall metrics similarly favored DenseNet‑121, particularly when expert segmentation was applied. Multivariate statistical analyses further confirmed significant performance differences across the evaluated architectures.
Collectively, these findings highlight the utility of incorporating information criteria into deep learning pipelines to enhance model robustness, generalizability, and interpretability in medical imaging applications, particularly for the non-invasive prediction of molecular biomarkers in glioblastoma.