48 Enabling paediatric optic pathway glioma segmentation via adult tumour transfer learning

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Abstract Introduction Accurate quantitative tumour assessment is central to longitudinal monitoring of paediatric optic pathway gliomas (OPGs), informing treatment decisions. However, MRI protocols in routine clinical practice are frequently incomplete due to patient tolerance or institutional variability. Moreover, since annotated paediatric datasets remain scarce, developing reliable automated tools has hitherto been infeasible. Here, we develop and evaluate a solution that addresses these challenges by leveraging robustness to MRI missingness and adult neuro-oncology transfer learning. Method A self-configuring nnU-Net framework was adapted using two integrated strategies. First, channel dropout was applied during training to simulate missing MRI sequences and encourage cross-modal learning, improving robustness to heterogeneous modality combinations encountered in clinical practice. Second, a two-phase transfer learning pipeline leveraged large-scale adult glioma data for feature pretraining (n = 1,251), followed by paediatric fine-tuning on a paediatric glioma cohort (n = 261). The approach was further evaluated on n = 17 independent paediatric OPG cases with heterogeneous MRI sequence combinations. Performance was assessed using the DICE similarity coefficient and percentage error in tumour volume (PEV). Results The adult-pretrained model demonstrated inconsistent segmentation (average DICE 0.256) and substantial volumetric overestimation. Following paediatric fine-tuning, segmentation accuracy improved. The fully fine-tuned model achieved an average DICE of 0.711. Volumetric bias was reduced, with an average PEV of 11.33%, compared with highly variable errors in the adult-only model (average PEV of 304.67%). Performance was stable across patterns of missing MRI sequences, reflecting successful sequence invariance and transfer learning. Conclusions Paediatric brain tumour segmentation is enhanced by leveraging data and technical advances from adult neuro-oncology. Our model supports quantitative monitoring of paediatric OPGs under heterogeneous clinical imaging conditions. Improved consistency of lesional estimation may facilitate more reliable longitudinal assessment in paediatric neuro-oncology practice.