Transfer learning enables automated pituitary gland segmentation on brain MRI in 6-year-old children

Pediatric pituitary gland (PG) segmentation is challenging owing to anatomical variability and the limited availability of labeled datasets, which restricts the direct application of adult-trained deep learning models. We aimed to develop and validate a specialized transfer learning (TL) pipeline for automated PG segmentation in 6-year-old children and to test two competing predictions, persistence versus catch-up, about the association of perinatal factors with PG volume at school age. A three-dimensional U-Net model pre-trained on 153 adult T1-weighted magnetic resonance images was adapted to a pediatric cohort of 61 children, using consensus ground truth established by majority agreement among three experts. The TL strategy incorporated dynamic region-of-interest (ROI) cropping derived from consensus centroids, ± 8-voxel spatial jitter, and extensive 20-fold data augmentation to improve model robustness. On an independent test set of 15 children evaluated with label-independent ROI placement, the TL model (Dice = 0.83 ± 0.23) significantly outperformed the adult-only model (Dice = 0.72 ± 0.22; P = 0.0002) in the evaluation that included one localization failure. Mean absolute relative volume error was 17.1% and the volume intraclass correlation coefficient was 0.48. Ablating the spatial-jitter component showed that it conferred robustness to ROI placement error rather than peak accuracy: with the ROI placed at the training-set mean centroid, the jitter-free variant fell to a Dice of 0.61 ± 0.38 whereas the full model reached 0.83 ± 0.23 ( P = 0.031). After adjusting for age at scan, the associations of gestational age and birth weight with PG volume were small (partial r = 0.17 and 0.18; 90% confidence interval [CI] − 0.02 to 0.36 and − 0.01 to 0.37), and equivalence testing excluded moderate-to-large residual effects (|partial r| ≥ 0.4; P = 0.021 and P = 0.025), a pattern more compatible with catch-up by school age, although small effects remain undetermined. These findings show that pediatric adaptation improves segmentation and suggest that moderate-to-large persistent perinatal associations with PG volume are unlikely in this cohort.