Comment on "Deep learning in CT-based organ-at-risk delineation for pediatric flank irradiation": Methodological and...

The study by Ding et al. on deep-learning-assisted organ-at-risk delineation for pediatric flank irradiation offers a valuable advancement toward automation in radiotherapy planning. However, several methodological and analytical gaps limit the confidence with which these findings can be generalized. Key issues include insufficient transparency of the manual annotation protocol and absence of baseline inter-observer variability metrics; possible circularity from using STAPLE consensus incorporating deep-learning contours; reliance solely on geometric similarity indices (Dice, HD95) without accompanying dosimetric validation; lack of uncertainty quantification or failure-mode analysis; and omission of workflow assessments beyond controlled settings. Together, these constraints obscure the true clinical impact of deep learning in radiotherapy contouring. Future investigations should prioritize transparent multicenter annotation standards, integrate probabilistic or uncertainty-aware models, include dosimetric endpoints, and evaluate performance within real-world clinical environments. Such measures will ensure that the promise of AI-assisted contouring translates into reproducible, safe, and clinically meaningful improvements in pediatric radiotherapy.