CT-based deep learning radiogenomics for predicting key glioma genotypes (IDH, ATRX, EGFR, TP53)

Area of Science:

• Neuroimaging
• Oncology
• Radiomics

Background:

• Molecular subtyping is crucial for glioma diagnosis and targeted therapy.
• Magnetic resonance imaging (MRI) is the current standard but can be time-consuming and contraindicated.
• Computed tomography (CT) is faster, more accessible, and suitable for emergency or resource-limited situations.

Purpose of the Study:

• To evaluate the accuracy of CT-based radiogenomic signatures combined with machine learning in predicting clinically relevant glioma molecular markers.
• To explore CT as a practical imaging modality for glioma molecular profiling.

Main Methods:

• Retrospective analysis of non-contrast CT (NCCT) scans from 197 adult gliomas.
• Development of predictive models for ATRX, EGFR, TP53, and IDH mutations using radiomic features and demographic data.
• Comparison of six classical machine-learning classifiers and deep-learning approaches (FCNN, TabNet) using ROC-AUC.

Main Results:

• Deep-learning methods significantly outperformed conventional classifiers for all molecular markers.
• TabNet achieved high ROC-AUCs for ATRX (0.900), TP53 (0.955), and EGFR (0.917).
• A custom fully connected neural network (FCNN) achieved a ROC-AUC of 0.971 for IDH prediction.

Conclusions:

• NCCT-based analytical methods accurately predict clinically relevant genetic mutations in gliomas.
• CT offers a practical alternative for molecular profiling in urgent or resource-limited settings.
• External validation is required for clinical translation, especially for preliminary EGFR findings in smaller subgroups.