P10.20.A RADIOMICS-BASED GRADING OF PAEDIATRIC GLIOMAS USING [18F] F-DOPA PET/CT: INSIGHTS FROM THE PERITUMORAL BORDER

Abstract BACKGROUND Gliomas, although rare, are the most common paediatric central nervous system tumours, accounting for 35% of cases and representing a major cause of cancer-related death in children. [18F] F-DOPA PET imaging plays a crucial role in their assessment. Differentiating high-grade (p-HGG) from low-grade gliomas (p-LGG) is essential for treatment planning, and non-invasive tools are particularly needed for lesions located in hard-to-reach areas. To this end, we developed a machine learning pipeline to assess whether radiomic features from PET imaging can enhance glioma grading, focusing on tumour borders due to the high prevalence (77%) of diffusely infiltrating lesions in our dataset. MATERIAL AND METHODS We retrospectively analysed [18F] F-DOPA PET scans of 103 paediatric patients diagnosed with p-HGG (n = 65) and p-LGG (n = 38) referred to Galliera Hospital between 2011 and 2024. Lesions were segmented using LIFEx, considering two volumes of interest (VOIs): (i) the whole tumour, defined as uptake >50% SUVmax; and (ii) a 5 mm peripheral margin. We extracted static metrics such as tumour-to-striatum (T/S) and tumour-to-background (T/N). We then computed 107 radiomic features via Pyradiomics and harmonised using Combat to address scanner variability. A nested cross-validation framework (5 outer, 10 inner folds) was adopted. We first trained the model with static features alone and then added radiomics from both VOIs. Feature selection used LASSO and Backward Feature Elimination; the selected features were employed to train a logistic regression model. RESULTS T/S was consistently selected in all folds, while T/N was excluded. From the tumour VOI, the most informative radiomics features were glcm_informational_measure_of_correlation, skewness, glcm_correlation, and elongation. Combining radiomic features with T/S improved classification performances: AUC (0.77±0.05 to 0.82±0.02), accuracy (0.67 ± 0.08 to 0.72 ± 0.04), recall (0.67±0.08 to 0.74±0.05), precision (0.71±0.07 to 0.72±0.04), F1 score (0.67±0.08 to 0.72±0.04) and MCC (0.37±0.15 to 0.43 ± 0.10). By Considering the tumour and margin VOIs, we achieved higher classification performances: AUC (0.84±0.03), accuracy (0.78±0.02), recall (0.78±0.02), precision (0.80±0.02), F1 score (0.78±0.02) and MCC (0.57±0.05). Key features included glcm_imc1, ngtdm_coarseness, and elongation. CONCLUSION Our preliminary study denotes the value of radiomics in differentiating p-HGG from p-LGG extracted from [18F] F-DOPA PET. Including radiomic metrics extracted from the tumour border improved classification performance, highlighting the relevance of peritumoral tissue in this paediatric cohort. Furthermore, this analysis supports the importance of computational and non-invasive methods to support clinical decision-making. Research supported by the project GLI-HOPE with grant number NET-2019-12371188.