Transferability of adult-trained CT organ segmentation to pediatric populations

Determine how the excellent segmentation performance achieved by state-of-the-art deep learning-based CT organ segmentation models trained on adult datasets translates to pediatric patients, specifically assessing the generalization of adult-trained models to pediatric CT images.

Background

The paper highlights that contemporary CT organ segmentation methods are typically trained on adult datasets and achieve strong performance in that domain. However, pediatric anatomy differs substantially from adult anatomy, raising concerns about whether adult-trained models will generalize to pediatric populations.

The authors frame this as an explicit unknown at the outset to motivate their paper, which evaluates TotalSegmentator on pediatric CT data and explores strategies such as data augmentation and continual learning to improve performance across age groups. While their experiments provide evidence for specific models and settings, the broader question of transferability across state-of-the-art adult-trained models is stated as unknown.