Abstract

Pediatric medical imaging presents unique challenges due to significant anatomical and developmental differences compared to adults. Direct application of segmentation models trained on adult data often yields suboptimal performance, particularly for small or rapidly evolving structures. To address these challenges, several strategies leveraging the nnU-Net framework have been proposed, differing along four key axes: (i) the fingerprint dataset (adult, pediatric, or a combination thereof) from which the Training Plan — including the network architecture—is derived; (ii) the Learning Set (adult, pediatric, or mixed), (iii) Data Augmentation parameters, and (iv) the Transfer learning method (fine-tuning versus continual learning). In this work, we introduce PSAT (Pediatric Segmentation Approaches via Adult Augmentations and Transfer learning), a systematic study that investigates the impact of these axes on segmentation performance. We benchmark the derived strategies on two pediatric CT datasets and compare them with state-of-the-art methods, including a commercial radiotherapy solution. PSAT highlights key pitfalls and provides actionable insights for improving pediatric segmentation. Our experiments reveal that a training plan based on an adult fingerprint dataset is misaligned with pediatric anatomy—resulting in significant performance degradation, especially when segmenting fine structures—and that continual learning strategies mitigate institutional shifts, thus enhancing generalization across diverse pediatric datasets. The code is available at https://github.com/ICANS-Strasbourg/PSAT.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/4351_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/ICANS-Strasbourg/PSAT

Link to the Dataset(s)

N/A

BibTex

@InProceedings{KirTri_PSAT_MICCAI2025,
        author = { Kirscher, Tristan and Faisan, Sylvain and Coubez, Xavier and Barrier, Loris and Meyer, Philippe},
        title = { { PSAT: Pediatric Segmentation Approaches via Adult Augmentations and Transfer Learning } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15966},
        month = {September},

}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper explores pediatric image segmentation via data augmentation and transfer learning strategies from adult imaging dataset. In particular, this work is based on the nn-UNet framework and the authors evaluate the best way to define: 1- training plan (network architecture and preprocessing), 2- training set, 3- data augmentation, 4- transfer learning or continual learning. The authors benchmark the performance on two pediatric CT datasets (one public, one private) and one adult CT dataset, and provide comparison with state-of-the-art methods.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    This paper addresses deep learning based pediatric image segmentation which remains challenging. Transfer learning and continuous learning appear to be simple solutions to tackle the scarcity of pediatric images.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    This paper does not present methodological contributions, but rather the authors benchmark different training strategies. It is not clear from the results reported which strategy is the best.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    Methodology:

    • The authors should consider training separate networks for each age group as the anatomy is very different between babies aged 0-2 years and children aged 14-16 years.

    Experiments:

    • The authors should report other validation metrics, such as Hausdorff distance, to provide a more complete assessment of the performance of the methods.

    Results:

    • Table 1 contains many structures of interest, reporting the average Dice would facilitate comparison between methods.
    • The authors should provide images with segmentation prediction to illustrate the performance of the different strategies.
    • It is not clear from Table 1 why PmSaAdTp is chosen as the best strategy.
  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (2) Reject — should be rejected, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The contributions of this paper are very limited.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The paper is suited for the Application Studies track.



Review #2

  • Please describe the contribution of the paper

    The paper presents a systematic analysis of strategies for adapting segmentation models to pediatric imaging tasks. Experiments compare models trained on adult, pediatric, and mixed datasets, exploring the effects of hyperparameter selection, data augmentation (including zoom in/zoom out), and transfer learning configurations (fine tuning vs. continuous learning). Results show that continuous learning can outperform standard fine tuning under domain shifts and that specific augmentations benefit pediatric datasets. The proposed models achieve competitive performance against industry benchmarks.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    1. Evaluation spans public adult, public pediatric, and internal pediatric datasets, with appropriate statistical testing.
    2. Novel analysis of zoom in and zoom out augmentations reveals clear benefits for pediatric data.
    3. Evidence supports continuous learning as an effective transfer learning strategy under domain shifts, offering valuable insights for both model development and deployment.
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    1. The manuscript lacks details on hyperparameter tuning for transfer learning experiments, a factor that could materially influence results.
    2. Continuous learning requires access to the entire adult pretraining dataset and may incur substantially higher computation compared to standard fine tuning; this trade off is not discussed.
    3. Even with continuous learning (PmSaAdTm), performance does not approach that of direct learning (PmSmAdTo) for most regions of interest, raising questions about its practical advantage.
  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
    1. Provide clear definitions of fine tuning and continuous learning early in the introduction.
    2. In Table 1, label each experimental run as either continuous learning (CL) or fine tuning (FT) for clarity.
    3. Consistently use the term “transfer learning” rather than “transfer,” including in the title.
    4. Revise the conclusion to emphasize that direct learning achieves the best performance when all relevant data are available.
  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — could be accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    This work is very important to optimize the strategies for adapting adult trained models or leverage adult datasets for pediatric segmentation applications. The details on learning rate and parameter tuning strategies used for the transfer learning experiments are critical and can influence the results. Recommend addressing these concerns before acceptance.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The authors have addressed the two major concerns I had and provided details in the manuscript on hyperparameters used and on the efficiency of CL vs FT vs direct training.



Review #3

  • Please describe the contribution of the paper

    The paper defines a systematic approach for optimizing transfer learning between adult and pediatric domains.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    • using multiple datasets
    • reporting results for multiple organs
    • proposing a very-well structured framework that is well-thought in both ML and clinical perspectives
    • presenting clear conclusions that can be used as a reliable reference for future studies
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    The study does not have any major weakness

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    As a reviewer, my number one goal is to help you improve your manuscript rather than judge if it should be published or not. I am aware of how difficult it can be to prepare a manuscript, and I wish you good luck with this publication. Please feel free to disagree with any comment that I add, as no one knows your research better than you. If you find my style of review helpful, please use it when you serve as a reviewer.

    I do not see any major issue in your manuscript. You have tackled a crucial issue and what you propose is a simple and robust methodology.

    [minor] there are platforms that help you anonymize your codes and submit them for review. It would be helpful to see your codes, but I understand the concerns when there is no preprint in place. Please ensure your codes are well documented and dockerize them if possible.

    [minor] Including a figure at the beginning of introduction is not common. I personally like out of the box ideas and do not see it as a negative point. However, it does not help establishing your story, as the differences between adult and pediatric data are recognized. You can move the figure to an appendix, but if you decide to keep it, please label the thee subplots with a,b, and c, and use those labels in the caption.

    [minor] Some parts of results could be included in methods. However, I do not want to create a low-return overhead for the authors and leave it to them to decide.

    With including lower-precision parameters and lower-resolution segmentation (downsampling) you would have a comprehensive study that would enable researchers to use the algorithms on cheaper computational resources. You may want to highlight this in the discussion for future works.

    Please ensure you do not redefine abbreviations over and over.

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (6) Strong Accept — must be accepted due to excellence

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The authors conduct a crucial study and provide a comprehensive and robust methodology with no major issue. I highly recommend accepting the manuscript.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The manuscript was already well beyond the acceptance threshold. Nonetheless, the authors responded to the review comments carefully and addressed the concerns as much as they could.




Author Feedback

We thank the reviewers for their valuable comments.

We agree with R1 that the paper does not present methodological novelty, we do not consider this a weakness. Our submission targets the Application Studies track, which welcomes work that adapts state-of-the-art methods to new contexts and demonstrates clinical relevance. Our study aligns with these goals. We tackle the challenging problem of deep learning–based pediatric image segmentation. Contrary to R1’s assertion of limited contributions, we introduce, as R2 highlights, a structured framework that is carefully designed from both machine learning and clinical perspectives. This framework enables a systematic comparison of adaptation strategies across key deployment scenarios (intra- vs inter-institutional), uncovering underexplored pitfalls and offering actionable insights that can serve as a reliable reference for future studies. The proposed strategies are validated on several datasets and outperform a leading commercial radiotherapy tool—evidence of clear translational value. We believe this work makes a meaningful contribution to the community.

R1 notes that it is unclear which strategy performs best. This reflects a misunderstanding: our study shows that no single approach outperforms others across all scenarios. Performance depends on the deployment context, which is why we explicitly separate results by scenario. As stated in the conclusion, fine-tuning (FT) yields the best performance in intra-institutional settings, while continual learning (CL) demonstrates stronger robustness in inter-institutional contexts.

R3 points out that our description of hyperparameter tuning lacks detail for transfer learning experiments. To address this, we revised the second paragraph of the “Training Procedures” section. To stay within the page limit, we removed Figure 1 (as suggested by R2). The updated paragraph reads:

For both FT and CL, we use the same “poly” learning rate decay schedule as in the pre-training phase, but we perform a grid search over initial learning rates in the interval [10^-3,10^-4] and over the number of epochs in the range [200,500], selecting the combination that yields the highest validation DSC. We use a smaller initial learning rate to preserve pretrained features, which is a common practice in transfer learning (e.g., Shin et al., 2016; Raghu et al., 2019). In the CL setting, we employ a rehearsal strategy as described by González et al. (2023) and Liu et al. (2024), and perform a grid search over the adult replay ratio in the range [0.25,1]. Overall, we observed that these hyperparameters had a limited impact on final performance, leading us to forego more complex hyperparameter optimization strategies.

R3 also notes that CL requires access to the full adult pretraining dataset and is more computationally intensive than FT, while its practical utility remains unclear given that direct learning performs slightly better. We addressed this by adding the following fourth point to the conclusion:

Direct learning seems to yield the highest performance; however, in inter-institutional scenarios, CL (PmSaAdTm) performs only marginally below direct learning (PmSmAdTo); similarly, in intra-institutional settings, FT (PmSaAdTp) closely matches direct learning. The practical advantage of transfer learning-based strategies lies in their efficiency: when a pretrained adult model is available, CL takes about 10 hours and FT about 2.5 hours, compared to about 25 hours for full retraining. Note that CL requires access to the full adult pretraining dataset and incurs higher computational cost than FT.

Finally, minor reviewer requests that are easy to implement and comply with MICCAI policy have been addressed in the camera-ready version.

Other comments from R1: Training separate models per age group is impractical given limited pediatric data. Hausdorff distances are in the code to be released and mirror DSC rankings, so were omitted from Table 1 for clarity.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    N/A

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Reject

  • Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



back to top