BIOM-81. IMAGING SURROGATE SIGNATURE OF THE GLIOMA EPIGENETIC LIQUID BIOPSY (GELB) SCORE FOR NON-INVASIVE DETECTION OF TUMOR PROGRESSION

Abstract BACKGROUND Gliomas are characterized by clinical and molecular heterogeneity and typically harbor a poor prognosis, particularly in high-grade forms. While multi-parametric Magnetic Resonance Imaging (mpMRI) assessment remains the clinical standard for monitoring disease progression, it is subjective and prone to misidentifying treatment-related changes as true progression. The Glioma epigenetic Liquid Biopsy (GeLB) score is a non-invasive blood-based biomarker capable of detecting tumor progression earlier than MRI. Here, we seek an objective radiomics-based surrogate signature of the GeLB score, towards democratizing GeLB’s detecting capability without the need for epigenetic analysis. METHODS We retrospectively analyzed mpMRI data (T1, T1ce, T2, T2-FLAIR) of 65 longitudinal paired GeLB-mpMRI samples (72.31% glioma vs 27.69% non-glioma), from 18 patients. GeLB over 50% defined glioma, and lower values non-glioma. 521 radiomic descriptors were extracted from i) the whole tumor (WT) defined by the abnormal T2-FLAIR envelope, and ii) the peritumoral edematous/infiltrated subregion (WT excluding enhancement and necrosis). A Support Vector Machine classifier was trained to distinguish GeLB-defined glioma/non-glioma, using sequential feature selection to eliminate redundant descriptors over 5-fold cross-validation. To mitigate overfitting and reduce selection bias associated with the limited sample size, performance metrics were averaged across five independent runs. RESULTS For edema-based descriptors, our signature yielded mean accuracy=67.69% (±0.102), sensitivity=80% (±0.163), and specificity=40% (±0.255). WT-based descriptors, slightly improved accuracy=73.85% (±0.078), and sensitivity=91.11% (±0.130), with comparable specificity=35% (±0.122). CONCLUSIONS Our findings support radiomic descriptors derived from clinical routine mpMRI sequences as promising imaging surrogates for the GeLB score, offering a potential non-invasive approach to improved progression detection, while obviating the need for epigenetic analysis. Although the specificity remains modest and the dataset limited in size, this proof-of-concept study highlights the potential of radiogenomics for improved response assessment. Future work will focus on validating these findings in larger, multi-institutional cohorts and leveraging more sophisticated descriptors.