Akademik Veri Yönetim Sistemi
48th ESNR Annual Meeting & 15th Asian-Oceanian Congress of Neuroradiology and Head and Neck Radiology & 34th Annual Meeting of the Turkish Society of Neuroradiology , İstanbul, Türkiye, 17 - 21 Eylül 2025, sa.9835261, ss.111, (Tam Metin Bildiri)
Artificial Intelligence-Enabled Radiogenomic Analysis for Non-invasive Molecular
Subtyping of Adult Gliomas Using Standard Computed Tomography
1 Zohal Alnour Ahmed Emam, 2 Emel Ada, 1 Berrin Çavuşoğlu, 3 Burçin Pehlivanoğlu, 4 M. Alper Selver, 5 Kadir
Akgüngör
1 Dokuz Eylul University, Institute of Health Sciences, Medical Physics Department, Izmir, Turkey
2 Dokuz Eylül University, Faculty of Medicine, Department of Radiology, Izmir, Turkey
3 Dokuz Eylul Hospital, Faculty of Medicine, Department of Pathology, Izmir, Turkey
4 Dokuz Eylul University, Electrical and Electronics Engineering Department and Izmir Health Technologies
Development and Accelerator (BioIzmir), İzmir, Turkey
5 Dokuz Eylul University, Department of Atomic and Molecular Physics, Izmir, Turkey
Abstract
Introduction: Molecular subtyping of gliomas is essential for treatment stratification according
to current WHO classification guidelines. Conventional molecular characterization requires
histopathological analysis following surgical resection or stereotactic biopsy, procedures
associated with perioperative morbidity and potential treatment delays. Radiogenomics, the
correlation of imaging phenotypes with genomic alterations, provides a non-invasive approach
for molecular profiling. This study investigates the application of machine learning algorithms to
standard computed tomography imaging for predicting key molecular markers in adult gliomas.
Methods: We retrospectively analyzed preoperative computed tomography scans from 197 adult
patients with histologically confirmed gliomas. Following quality control and data exclusion
criteria, the final cohort comprised 81 cases for alpha-thalassemia/mental retardation X-linked
(ATRX) analysis, 17 cases for epidermal growth factor receptor (EGFR) analysis, 71 cases for
p53 protein status analysis, and 183 cases for isocitrate dehydrogenase (IDH) analysis. Radiomic
feature extraction yielded 199 quantitative imaging biomarkers. Feature selection was performed
using hybrid Least Absolute Shrinkage and Selection Operator-Recursive Feature Elimination
(LASSO-RFE) and Gradient Boosting-Recursive Feature Elimination (GB-RFE) methods. Eight
machine learning classifiers were evaluated: Support Vector Machines (SVM), Random Forest
(RF), Multi-Layer Perceptron (MLP), Logistic Regression (LR), Extreme Gradient Boosting
(XGBoost), k-Nearest Neighbors (k-NN), TabNet, and deep neural networks (DNN) for
prediction of IDH, p53 protein status, EGFR, and ATRX mutations.
Results: TabNet performed with an area under the receiver operating characteristic curve (AUC)
of 0.900 and precision-recall area under curve (PR-AUC) of 0.983 for ATRX mutation
prediction; AUC of 0.917 and PR-AUC of 0.950 for EGFR amplification; and AUC of 0.955 and
PR-AUC of 0.917 for p53 protein status. Deep neural networks achieved AUC of 0.971 and PR-
AUC of 0.904 for IDH mutation status. Cross-validation analysis yielded coefficient of variation
(CV) values of 0.5-2% for ATRX (AUC and PR-AUC) and 0% for EGFR (both metrics).
Discussion & Conclusion: This study demonstrates that CT-based radiogenomic analysis can
successfully identify molecular markers in adult gliomas using machine learning approaches.
Through inclusive evaluation of eight algorithms, TabNet and deep neural networks achieved
optimal performance across four genetic targets. Cross-validation results confirmed model
stability and reproducibility. This work addresses the limited research utilizing CT for glioma
radiogenomics and establishes a validated methodology for non-invasive molecular profiling
using widely available imaging technology. The findings provide a foundation for expanding
radiogenomic applications beyond advanced MRI-based approaches.
Keywords: Radiogenomics; Artificial Intelligence; Glioma; Molecular Profiling; Computed
Tomography
Abstract tables:
Table 1 . Performance comparison of machine learning models for molecular marker prediction showing area
under the receiver operating characteristic curve (AUC) and precision-recall area under the curve (PR-AUC) for
each genetic target
Gene Model AUC PR
AUC
ATRX KNN 0.593 0.871
ATRX LR 0.579 0.903
ATRX MLP 0.586 0.903
ATRX RF 0.739 0.952
ATRX SVM 0.671 0.932
ATRX TabNet_model 0.900 0.983
ATRX XGBoost 0.525 0.880
ATRX nn_model 0.691 0.938
EGF
R
KNN 0.750 0.768
EGF
R
LR 0.583 0.793
EGF
R
MLP 0.583 0.793
EGF
R
RF 0.750 0.854
EGF
R
SVM 0.750 0.893
EGF
R
TabNet_model 0.917 0.950
EGF
R
XGBoost 0.458 0.554
EGF
R
nn_model 0.667 0.830
p53 KNN 0.511 0.463
p53 LR 0.546 0.487
p53 MLP 0.568 0.505
p53 RF 0.261 0.221
p53 SVM 0.591 0.392
p53 TabNet_model 0.955 0.917
p53 XGBoost 0.432 0.282
p53 nn_model 0.660 0.658
IDH KNN 0.929 0.640
IDH LR 0.967 0.886
IDH MLP 0.914 0.795
IDH RF 0.924 0.651
IDH SVM 0.900 0.789
IDH TabNet_model 0.890 0.754
IDH XGBoost 0.910 0.620
IDH nn_model 0.971 0.904
Table 2 . Five-fold cross-validation stability analysis coefficient of
variation (CV) of model performance metrics (AUC and PR-
AUC) across different radiomic feature sets for molecular
marker prediction
Gene Metric Mean Std
Dev
CV
(%)
ATRX AUC 0.966 0.020 2.10
ATRX PR
AUC
0.994 0.005 0.46
p53 AUC 0.960 0.029 2.97
p53 PR
AUC
0.906 0.074 8.15
EGF
R
AUC 1.000
0
0.000 0
EGF
R
PR
AUC
1.000
0
0.000 0
IDH AUC 0.823 0.135 16.40
IDH PR
AUC
0.679 0.124 18.33
Abstract images:
Figure 1 . Receiver operating characteristic (ROC) curve for ATRX mutation prediction using the TabNet model.
The curve demonstrates the trade-off between sensitivity and specificity, with an area under the curve (AUC) of
0.900, indicating high discriminative performance for ATRX mutation status classification in adult gliomas based
on computed tomography radiomic features.
Figure 2 . Receiver operating characteristic (ROC) curve for EGFR amplification prediction using the TabNet
model. The analysis reveals robust diagnostic capability with an area under the curve (AUC) of 0.917,
establishing strong predictive accuracy for EGFR amplification detection in glioma cases through CT-based
radiomic feature analysis.
Figure 3 . Receiver operating characteristic (ROC) curve for p53 protein status prediction using the TabNet
model. The performance evaluation achieved an area under the curve (AUC) of 0.955, reflecting exceptional
classification capability for p53 protein status determination using computed tomography-derived quantitative
imaging parameters.
Figure 4 . Receiver operating characteristic (ROC) curve for IDH mutation prediction using the deep neural
network (DNN) model. The analysis achieved superior diagnostic performance with an area under the curve
(AUC) of 0.971, establishing the most accurate molecular classification among all genetic targets through CT-
based radiogenomic profiling.