Integrated CT-PET radiogenomics and graph-based multi-task learning for preoperative prediction of key glioma
Aisha Abdallah1, Basmah Altuwayjiri2, Abdullah M K Albloshi3
1Department of Health Information Management and Technology (HIMT), College of medical and applied sciences, University of Hafr Al-Batin, Hafar Al-Batin, Saudi Arabia.
Background:
The clinical management of glioma is increasingly dependent on the tumor's molecular profile, particularly the mutation status of Isocitrate Dehydrogenase (IDH), the promoter methylation of O-6-Methylguanine-DNA Methyltransferase (MGMT), and the co-deletion of chromosomal arms 1p and 19q. Determining these markers requires invasive tissue sampling. This study aimed to develop and validate a multimodal artificial intelligence (AI) framework for the non-invasive and simultaneous prediction of these three key markers using routine preoperative CT and PET imaging.
Methods:
In this multi-center, retrospective study, data from 1,472 glioma patients with confirmed histopathological diagnosis and molecular profiles were collected from four international centers. CT and PET images were fused using a Wavelet Transform. Subsequently, a comprehensive feature set was extracted, including 24 clinical variables, 3,654 radiomics features, and 43,776 deep features. The deep features were extracted from three advanced architectures: EfficientNet-B7, Swin Transformer, and DINOv3. The Boruta algorithm was used for dimensionality reduction and optimal feature selection. Finally, a Multi-Task Learning model based on a Graph Neural Network (GNN) was designed for the simultaneous prediction of IDH, MGMT, and 1p/19q status. The model's performance was evaluated using 5-fold cross-validation on the training dataset (n = 1,158) and finally validated on a completely independent external test set (n = 314).
Results:
The proposed multi-task GNN model is expected to achieve outstanding performance on the external test set: predicting IDH mutation status with an Area Under the Curve (AUC) exceeding 0.96, MGMT methylation with an AUC above 0.93, and 1p/19q co-deletion with an AUC greater than 0.95. This integrated approach is anticipated to significantly outperform models that predict each marker individually. Interpretability analysis using SHAP will reveal the combined role of clinical features (e.g., age), radiomics features (e.g., texture matrices), and deep features (e.g., abstract patterns from DINOv3) in predicting each molecular marker.
Conclusion:
This study introduces a powerful and integrated AI framework capable of predicting the vital molecular profile of glioma with very high accuracy and non-invasively. This tool has the potential to transform diagnostics in neuro-oncology and facilitate personalized therapeutic decision-making before surgery.
