Decoding glioma molecular heterogeneity by advanced radiomic and radiogenomic analyses

High-grade gliomas (HGGs) are the most common primary brain malignancy. One of their main characteristics is a great inter and intra-tumor molecular, cellular, and histological heterogeneity. In particular, the molecular characterization of HGGs has gained importance in neuro-oncology, as it enables more precise prognostic stratification than histology alone. However, HGGs diagnosis still relies on histopathological analyses of surgical specimens, which do not fully represent the entire tumor. This highlights the need for reliable, non-invasive biomarkers for comprehensive disease assessment and dynamic monitoring. Magnetic resonance imaging (MRI), including conventional (cMRI) and advanced (aMRI) sequences, provides a macroscopic perspective on HGG heterogeneity, serving as an imaging biomarker. Quantitative features derived from cMRI and aMRI can be integrated in radiomic and radiogenomic models to uncover distinct sub-visual imaging patterns with clinical and molecular relevance. The aim of this thesis was to decode glioma heterogeneity by advanced radiomic and radiogenomic analyses through three experiments. In the first experiment, I applied a novel habitat imaging approach to characterize HGGs heterogeneity, integrating information on Hypoxia, advanced PERfusion, and DIffusion MRI techniques together with amino acid PET in a unique HYPERDIrect map. I identified a robust and reproducible pattern of habitats’ distribution and a strong correlation between habitats’ predicted characteristics and histological, immunohistochemical, and molecular tissue properties. Furthermore, I evaluated the potential use of the HYPERDIrect map as a clinical predictor, finding an interesting correlation between habitats’ volumes and patient prognosis. In the second experiment, I conducted a structured analysis of the interplay between MRI-derived metrics and circulating cell-free DNA (ccfDNA) levels and their potential integration into machine-learning models to predict patient prognosis. I observed a weak correlation between these two variables at single time points but a strong relationship in longitudinal analyses, highlighting the potential use of ccfDNA for HGGs follow-up. Interestingly, radiomic features improved the prediction of ccfDNA levels compared to tumor volume alone, suggesting that factors related to tumor heterogeneity significantly influence the amount of tumor-related cfDNA released in the bloodstream. Finally, the integration of radiomic features and ccfDNA outperformed ccfDNA alone in predicting progression-free survival. In the third experiment, I investigated how subvisual quantitative imaging features of gliomas relate to their molecular status. I defined a robust pipeline to develop a radiogenomic model capable of correctly classifying lesions according to IDH-molecular status with high accuracy, specificity, and sensitivity. Additionally, I evaluated the radiological significance of the extracted radiomic features. Finally, I developed a radiomic model incorporating the same features along with clinical variables, which demonstrates strong performance in identifying patients with a worse prognosis. In conclusion, these experiments show different imaging biomarkers directly related to the heterogeneity, molecular features, and clinical behavior of gliomas. The findings of this thesis highlight the significant potential of non-invasive imaging biomarkers, alone or in combination with other biological markers, to provide a comprehensive, tumor-wide, and dynamic characterization of gliomas, making them highly valuable tools for patient stratification and clinical evaluation.