Paediatric thoracal and abdominal organ segmentation in CT and MRI using global intensity non-linear augmentation

Objectives Robust tools for automated segmentation of paediatric organs in cross-sectional imaging remain limited, particularly compared to well-established models for adult populations such as TotalSegmentator. The objective of this study was to develop a segmentation model and incorporate Global Intensity Non-Linear (GIN) augmentation to determine its potential for robust organ segmentation in paediatric computed tomography (CT) and magnetic resonance imaging (MRI). Methods We trained a 3D U-Net model on the public Pediatric-CT-SEG dataset, incorporating GIN to facilitate transfer from CT to MR images. The model was evaluated on a subset of the public CT dataset (n = 70) and fully independent paediatric MRI examinations (n = 42). We compared our novel GIN model with a structurally identical baseline model without GIN as well as TotalSegmentator using Dice similarity coefficient (DSC) and normalized surface distance (NSD). Results The baseline and GIN model achieved similar mean DSC and NSD values for CT examinations across all organs: DSC 0.91 ± 0.13 (baseline) and 0.91 ± 0.12 (GIN), NSD 0.77 ± 0.16 (baseline) and 0.78 ± 0.15 (GIN). TotalSegmentator achieved significantly lower overall DSC than the GIN model of 0.88 ± 0.12 and NSD of 0.71 ± 0.19, p < 0.001. For MRIs, our GIN model showed significantly higher overall DSC (0.72 ± 0.24) and NSD (0.46 ± 0.17) compared to the baseline model (DSC 0.18 ± 0.26; NSD 0.11 ± 0.16) and to TotalSegmentator (DSC 0.56 ± 0.37; NSD 0.30 ± 0.26), resulting in p < 0.001 for all organs except lungs. Conclusion Our findings highlight the potential of deep-learning models with GIN augmentation to enable robust organ segmentation, and consequently reliable organ volumetry, in paediatric CT and MRI examinations within clinical workflows. To our knowledge, the proposed GIN model is the first to enable robust multi-organ segmentation in paediatric MRI and CT examinations.