Performance of automated segmentation models for the delineation of paediatric craniopharyngiomas.

Motivation: Accurate automated segmentation of paediatric craniopharyngiomas could improve and standardise measurements of tumour size used to assess response. Goal(s): To assess the performance of published, freely available model architectures for automated whole-tumour segmentation of paediatric craniopharyngiomas on routine brain MRI. Approach: Ground-truth segmentations were compared to predicted masks performed with 3 different models (1 DeepMedic and 2 nnU-Net based), using measurements of the Dice coefficient and percentage error in tumour volume. Results: For whole-tumour segmentation of paediatric craniopharyngioma, the highest Dice scores alongside accurate volume measurements were achieved with models published by Ruffle et.al.(2023). Impact: Whole-tumour segmentations using models published by Ruffle et.al.(2023) have the potential to aid with volume measurements of paediatric craniopharyngioma on routine brain MRI. Clinical applications would benefit from separate segmentations of solid and cystic tumour in future analysis.