Deep learning-based pediatric swallowing video segmentation from sparse multi-annotator endoscopic data: comparing supervised U-Net models and in-context segmentation

Accurate analysis of pediatric Fiberoptic Endoscopic Evaluation of Swallowing (FEES) is essential for detecting swallowing abnormalities. Automated segmentation of FEES data remains challenging due to high inter-expert annotation variability, severe class imbalance, and the limited availability of labeled pediatric datasets. This study presents a pediatric FEES segmentation framework and a comparative evaluation of supervised and training-free approaches using sparse annotations from five expert clinicians. We evaluated class-wise U-Net + + models, the recently proposed VM-UNet architecture, and Temporal, a pretrained video object segmentation framework applied in a training-free, in-context learning setting. Six anatomical and pathological classes were considered: laryngeal inlet, vocal cords, penetration, residue, spillage, and salivary pooling. Two complementary experiments were performed: an internal frame-level five-fold cross-validation on 38 videos, and an independent patient-level evaluation in which models trained on the 38 videos were tested on 50 newly annotated videos from different patients. In the internal cross-validation, mean Dice was 0.70, 0.69, and 0.77 for U-Net + + , VM-UNet, and Temporal, respectively. On the independent test set, mean video-level Dice across five classes was 0.45, 0.44, and 0.40 (laryngeal inlet Dice up to 0.67 and presence score up to 0.96). These findings show that models trained on sparse, mostly single-expert annotations can segment anatomical structures in unseen pediatric patients. Supervised models were more accurate for anatomical structures, whereas all approaches performed comparably for swallowing events. Penetration remained the most challenging class, consistent with its lowest inter-expert agreement. Frame-level validation substantially overestimated performance, particularly for retrieval-based in-context models.