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

Abstract

Purpose: Molecular subtyping guides diagnosis and targeted therapy for gliomas. Although MRI-the current imaging standard-can be time-consuming and is sometimes contraindicated, computed tomography (CT) is faster, more widely available, and often preferable in emergency and resource-limited settings. We evaluated whether CT-based radiogenomic signatures combined with machine learning could accurately predict clinically relevant glioma molecular markers.

Methods: In this retrospective study, we characterised non-contrast CT (NCCT) scans from 197 adults with histologically confirmed gliomas. Models were developed to predict mutations in four molecular markers: ATRX (n=81), EGFR (n=17), TP53 (n=71), and IDH (n=183). We extracted 208 quantitative radiomic features and added basic demographic variables. Feature selection used LASSO-RFE and Gradient Boosting-RFE with cross-validation. Six classical machine-learning classifiers and deep-learning approaches - including custom fully connected neural networks (FCNN) and TabNet - were trained and compared using ROC-AUC as the primary performance metric.

Results: Deep-learning methods outperformed conventional classifiers for all targets. TabNet achieved ROC-AUCs of 0.900 (95% CI: 0.717-0.989; ATRX), 0.955 (95% CI: 0.661-0.978; TP53), and 0.917 (95% CI: 0.858-0.975; EGFR). A custom FCNN obtained a ROC-AUC of 0.971 (95% CI: 0.876-0.995) for IDH. Cross-validation coefficients of variation were 2.0% for ATRX and 3.0% for TP53 and EGFR, and 16.0% for IDH. Deep-learning approaches yielded statistically significant improvements over conventional methods (p-values ranging from <0.05 to <0.001).

Conclusion: NCCT-based analytical methods were able to predict clinically relevant genetic mutations in gliomas and demonstrated performance comparable to established techniques. These findings suggest that CT may serve as a practical option for molecular profiling in urgent or resource-limited settings. Nonetheless, external validation is necessary prior to clinical translation. The EGFR findings, arising from a small EGFR-tested subgroup (n = 17), remain preliminary; the observed performance may not reflect generalisable accuracy, and independent validation in larger cohorts is required.