Comment on “Deep learning in CT-based organ-at-risk delineation for pediatric flank irradiation”: Methodological and clinical considerations

Abstract

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.

Dear Editor,

We read with great interest the recent article by Ding et al. that evaluated the role of deep learning (DL) in CT-based organ-at-risk (OAR) delineation for pediatric flank irradiation [1]. The authors reported significant reductions in contouring time, improved accuracy, and decreased inter-observer variability (IOV). While this represents an important contribution to the emerging field of AI-assisted radiotherapy in children, several methodological and analytical issues remain that may limit the robustness and generalizability of these findings.

First, annotation protocol transparency and inter-observer variability baseline: the study reports IOV reductions but does not describe the annotation protocol in sufficient detail or provide baseline IOV estimates for manual-only contours across centers [2]. Without this information, it is difficult to disentangle the true DL-related benefits from the natural variability in humans. A clearly defined manual contouring protocol and baseline metrics are necessary for reproducibility and a fair assessment.

Second, STAPLE circularity bias: The STAPLE consensus method used as the reference incorporates DL-revised contours, which risks circularity and may artificially inflate similarity with the DL arm. Independent multi-expert references are recommended to avoid circularity bias [3]. This could otherwise overstate the model’s true impact on reducing the inter-observer variability.

Third, the limitations of the Dice/HD95 metrics without dosimetric validation – The evaluation relies only on geometric similarity indices (Dice, HD95, MSD). However, small geometric differences can cause significant dosimetric variations in radiotherapy. Without dose–volume or plan quality analysis, the actual clinical relevance of these improvements remains uncertain [4].

Fourth, lack of uncertainty estimation and failure-mode analysis: The study does not include any uncertainty quantification or failure-mode analysis, leaving clinicians without guidance on when the model might fail and require manual verification. Models that provide probabilistic segmentations or voxel-level variance maps have been proposed, especially in radiotherapy, to alert users to potentially unreliable contours [5]. Incorporating such uncertainty measures can strengthen the model interpretability and clinical safety.

Fifth, generalizability to real-world workflows: Although the reported time savings come from a controlled workshop setting, integration challenges that normally occur in daily clinical workflows (e.g., software loading times, user correction burden, and system latency) were not assessed. Prior reports have shown that such overheads can diminish the apparent efficiency gains observed in experimental settings [6,7]. Hence, these results may represent an optimistic estimate compared to those obtained in routine clinical practice.

In conclusion, while Ding et al. presented promising evidence that deep learning can enhance and standardize OAR contouring in pediatric flank irradiation, several gaps limit the confidence with which these findings can be applied in real-world practice. The lack of annotation protocol transparency, potential STAPLE circularity, exclusive use of geometric metrics without dosimetric validation, absence of uncertainty analysis, and controlled settings together restrict the clinical interpretability of the results. Addressing these gaps through multi-institutional benchmarking, uncertainty quantification, dosimetric endpoint inclusion, and real-world workflow evaluation is crucial to establish the true clinical utility and safety of DL-assisted contouring in pediatric radiotherapy [[5], [6], [7]].

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Declaration of generative AI and AI-assisted technologies in the manuscript preparation process:

During the preparation of this work the authors used ChatGPT in order to improve readability. After using this tool the authors reviewed and edited the content as needed and take full responsibility for the content of the published article.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.