80 The Multi-Modal Combination of Magnetic Resonance Spectroscopy and Apparent Diffusion Coefficient Features Improves Radiological Prognostic Stratification of Paediatric Medulloblastoma

Abstract Introduction Magnetic Resonance Spectroscopy (MRS) and Apparent Coefficient Maps (ADC) features have previously demonstrated prognostic value, identifying high-risk groups of paediatric medulloblastoma (MB) and predicting patient survival. Despite evidence in diagnostic applications that their multi-modal combination can provide further added value over either modality individually, the combination of ADC and MRS has not been evaluated for MB prognostics. Method This study utilises the MRS and ADC from 44 MB patients recruited into the Imaging of Tumours Study (CNS-2004-10). 107 radiomic features (intensity, shape-based, textural) were extracted from ADC maps via PyRadiomics. 38 MRS metabolite concentrations were estimated by TARQUIN. Each feature set was reduced separately, removing redundant/unstable features. Random Forests were used to predict overall survival (OS) at 1-, 2- and 5-years, evaluating balanced accuracy and ROC-AUC. Models were assessed for each set of features individually, and then combined. Results The models using ADC, MRS, and both ADC & MRS features each achieved balanced accuracy of 0.63, 0.67, and 0.75 in the 2-year OS task, with ROC-AUC 0.69, 0.73, 0.820 respectively. Similar trends were identified for 1- and 5-year OS. A significant (p < 0.05) hazard-ratio of 7.31 existed between the multi-modal model’s predicted groups, with no significant finding for single-modality models. Low Total N-acetyl-aspartate, high Total Lipids and Macromolecules (1.3ppm), low Tumour Diameter, and low ADC 10th Percentile were the strongest indicators of poor survival identified by the multi-modal model. Conclusions This novel combination of ADC and MRS features demonstrates improved risk stratification and OS classification over either ADC or MRS features individually, at all timepoints. Subsequent work will include survival analyses to estimate patient risk over time, and integrate known clinical, demographic and histological features to robustly evaluate added value.