Detection of EGFR gene mutations in glioblastoma: Utilizing information complexity in developing AI-based decision support system.
Authors
Affiliations (3)
Affiliations (3)
- Koç University, College of Engineering, Computer Engineering Department, Turkey.
- Mimar Sinan Fine Arts University, Faculty of Science and Letters, Department of Statistics, Turkey. Electronic address: [email protected].
- University of California, Berkeley, Department of Electrical Engineering and Computer Sciences, USA.
Abstract
Glioblastoma is the most common and deadly brain cancer, known for its rapid progression and heterogeneity at microscopic and macroscopic levels. This heterogeneity is influenced by factors such as tumor cell density, involvement of normal tissue, and gene expression profiles. Mutations in EGFR gene are associated with shorter recurrence intervals and poorer survival outcomes in GBM patients. Non-invasive imaging techniques like MRI can provide valuable insights into EGFR mutations. To reduce the risks of brain biopsies and sampling errors, this study introduces an AI-based decision support system (DSS) for classifying EGFR mutations in GBM patients through automated segmentation of tumorous regions using MRI. The DSS employs deep neural networks (Inception ResNet-v2, DenseNet-121, and ResNet-50) trained on a GBM dataset from Memorial Hospital in Istanbul, which includes three MRI input types: expert segmented, without segmentation, and without tumor. Information criteria (IC) were used to guide model selection by balancing predictive performance and structural complexity. DenseNet-121 showed superior performance, with accuracy scores of 0.952, 0.942, and 0.938 for expert segmented, without segmentation, and absence of tumor inputs, respectively. Precision and recall metrics were also highest for DenseNet-121, especially with expert-segmented inputs. A multivariate statistical analysis confirmed significant differences across model performances. The results underscore the value of integrating information criteria into deep learning pipelines to enhance model robustness and interpretability in medical imaging applications.