Leveraging Adult Brain Tumour AI Models to Predict Paediatric Neurooncology Survival

Abstract AIMS Accurate, automated, segmentation of brain tumours for paediatric patients can significantly impact diagnosis, treatment planning, and prognosis. However, it remains challenging due to data scarcity, tumour heterogeneity and class imbalances introducing bias. A systematic review of ML-based models using MRI for tumour classifi- cation for pathological type and lesion segmentation found 29/49 studies included paediatric datasets, with just four leveraging adult-trained models. This study assesses the generalisability of an AI-based algorithm trained on over 5000 adult neuro-oncology MRIs for paediatric (i) brain tumour segmentation (ii) survival prediction. METHODS Using a validated adult glioma deep learning segmentation model (Ruffle et al., Brain Comms 2023), we quan- tified its out-of-distribution fidelity in our large dataset (285 patients, 346 images, mean age 11, IQR 6-17, 197 male) of paediatric brain tumours. Our cohort includes nine pathological types, most commonly Medulloblas- toma (53%), GBM (42%), Germinoma/GCT and DIPG (both 23%). The quality of the segmentations was assessed qualitatively by an expert paediatric neuroradiologist and 20 images were manually segmented to enable quan- titative scoring. Imaging was registered to the MNI152 1mm3 isotropic template, with disconnectome maps generated from the enhanced and non-enhanced components (Thiebaut de Schotten et al., Nat Comms 2020). Using a validated autoencoder (Ruffle et al., BNOS 2025 #163), we derived latent features from 189 lesions and trained a random forest survival model. RESULTS The adult segmentation model produced qualitatively high-quality results for 259/346 cases (75%) with a median Dice of 0.80. The survival model achieved 79% accuracy, 67% precision, 62% sensitivity, and 83% specificity in predicting mortality at the last contact point for the holdout set. CONCLUSION This study demonstrates the potential generalisability of adult-trained deep-learning models for paediatric neuro-oncology patients across tumour segmentation and survival prediction. Future work will refine these methods with larger cohorts and transfer learning.