Pith. sign in

REVIEW 5 major objections 5 minor 27 references

A Large Convolutional Neural Network for Clinical Target and Multi-organ Segmentation in Gynecologic Brachytherapy with Multi-stage Learning

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read GynBTNet, a 440-million-parameter CNN trained in three stages, outperforms nnU-Net and Swin-UNETR on gynecologic brachytherapy CT segmentation.

desk verdict Useful multi-stage pretraining recipe for GYN-BT segmentation, but the pretraining benefit is not significant on primary metrics and the paper overclaims. read the letter →

arxiv 2506.01073 v1 pith:YSWQ4NXK submitted 2025-06-01 cs.CV

Mingzhe Hu , Yuan Gao , Yuheng Li , Ricahrd LJ Qiu , Chih-Wei Chang , Keyur D. Shah , Priyanka Kapoor , Beth Bradshaw
show 5 more authors
Yuan Shao Justin Roper Jill Remick Zhen Tian Xiaofeng Yang
This is my paper
classification cs.CV
keywords gynecologicbrachytherapyclinicaltargetvolumesegmentationorgan-at-riskself-supervisedpretrainingsparsesubmanifoldconvolutionmulti-stagelearningCTimagedeepconvolutionalnetwork
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

GynBTNet is a 440-million-parameter convolutional network that segments the high-risk clinical target volume and four organs at risk in gynecologic brachytherapy CT scans. The paper's central claim is that a three-stage schedule—self-supervised masked-image pretraining on 6,157 unlabeled CT scans, supervised fine-tuning on a 104-structure corpus, then task-specific fine-tuning on 145 brachytherapy cases—lets a large CNN beat two strong baselines, nnU-Net and Swin-UNETR, despite the scarcity of annotated brachytherapy data. On an independent 29-case test split, GynBTNet reaches Dice coefficients of 0.837 for the clinical target volume, 0.940 for the bladder, 0.842 for the rectum, and 0.871 for the uterus, with lower boundary distances than the baselines. Self-supervised pretraining is the load-bearing ingredient: removing it drops CTV Dice by 0.08, the largest gap of any structure.

What carries the argument

The load-bearing mechanism is the three-stage training pipeline with sparse submanifold convolution in its self-supervised stage. Masked image modeling randomly occludes non-overlapping cubic patches of size 7×7×8 at 60 percent probability after the encoder's final downsampling, and sparse submanifold convolution restricts all computation to unmasked voxels, preserving spatial structure and avoiding feature dilution from masked regions. The encoder weights learned this way initialize a dense five-stage residual encoder-decoder whose depth, width, and compound scaling follow nnU-Net-style automatic configuration and EfficientNet-style scaling, reaching about 440 million parameters; the decoder is reinitialized at supervised fine-tuning.

What would settle it

Run GynBTNet and nnU-Net on an external multi-institution gynecologic brachytherapy dataset with contours from multiple oncologists and a mix of applicator types; if the CTV Dice gap between the models disappears or reverses, or the boundary-distance gains do not persist, the claimed superiority of the multi-stage pretraining strategy would be falsified.

Watch

Extended reading notes

Core claim

The paper establishes that a convolutional segmentation model can be pretrained like a foundation model and then specialized to a low-data clinical task with measurable gains. GynBTNet learns general anatomy by masked image modeling on heterogeneous CT scans, using sparse submanifold convolutions so that only unmasked voxels participate in reconstruction; the dense encoder is then fine-tuned on TotalSegmentator and finally on a dedicated gynecologic brachytherapy dataset. The authors report that this multi-stage model significantly outperforms nnU-Net and Swin-UNETR on all structures except the sigmoid colon, where all models plateau near a Dice coefficient of 0.60, and that the pretraining stage contributes consistent gains, most strongly for the clinical target volume.

Load-bearing premise

The evaluation assumes that 29 test cases from 30 patients at one institution, contoured by one oncologist and reviewed by a senior oncologist, are representative enough to show that GynBTNet's advantages over nnU-Net and Swin-UNETR hold for the wider gynecologic brachytherapy population.

Editorial extensions

If this is right

  • Pretraining on unlabeled CT scans is worth about +0.08 Dice for the clinical target volume, so centers adopting GynBTNet should keep the self-supervised stage rather than train from scratch.
  • For bladder, rectum, and uterus, GynBTNet's boundary metrics (HD95 and average surface distance) are consistently lower than both baselines, meaning auto-contours require less manual correction on these structures.
  • The sigmoid colon remains a failure mode at DSC 0.603, so clinical workflows should still flag sigmoid contours for review.
  • Because the three-stage recipe uses nnU-Net-style automated configuration and public pretraining corpora, the same schedule can be applied to other low-data segmentation tasks without hand-tuning.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper does not test how little task-specific data GynBTNet needs; a natural extension is to fine-tune the pretrained model on 10, 20, and 40 brachytherapy cases to map the data-efficiency curve.
  • The sparse-submanifold masked-modeling trick is architecture-agnostic: any dense CNN encoder could adopt it, so the gain it produces on boundary-heavy structures such as the CTV is a testable mechanism for other CT segmentation tasks.
  • A direct extension suggested by the paper's own discussion is CT-plus-MRI training, likely to matter most for the sigmoid colon, where all models plateau near Dice 0.60 because CT contrast is insufficient.
  • If the 0.08 CTV gain transfers across institutions, multi-stage pretraining could shrink annotation requirements for brachytherapy programs, since one oncologist contouring a few dozen local cases may suffice for fine-tuning.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper introduces GynBTNet, a large convolutional network for segmenting the high-risk clinical target volume (HRCTV) and organs at risk (OARs) in gynecologic brachytherapy CT images. The proposed method uses a three-stage learning strategy: self-supervised masked image modeling pretraining on 6,157 unlabeled CT scans using sparse submanifold convolution, supervised fine-tuning on the TotalSegmentator dataset with 1,204 annotated CT scans, and task-specific fine-tuning on an institutional dataset of 145 cases from 30 patients. The model is compared with nnU-Net and Swin-UNETR on 29 held-out test cases, reporting DSC, HD95, and ASD. The authors report that GynBTNet outperforms both baselines on most structures, that self-supervised pretraining yields consistent improvements, and that the gains are statistically significant. The paper also includes an ablation comparing GynBTNet with and without pretraining.

Significance. If the performance claims are robust, the multi-stage pretraining pipeline could be a valuable contribution to GYN brachytherapy segmentation, where annotated data are scarce. The paper has several strengths: it uses a patient-level split to avoid data leakage, leverages large public CT datasets for pretraining, evaluates both overlap and surface-based metrics, and reports statistical tests. The focus on a clinically important and relatively understudied application is commendable. However, the current evidence is weakened by a small single-institution test set, missing ablations, and statistically unsupported overclaims about the consistency and significance of the improvements. The central idea is plausible and the paper provides a useful empirical starting point, but the conclusions need to be tempered and the analyses strengthened before publication.

major comments (5)
  1. [Section 3.2, Table 1] The claim that 'GynBTNet consistently outperforms both nnU-Net and Swin-UNETR across all organs' is contradicted by the sigmoid colon DSC results: GynBTNet achieves 0.603 ± 0.075, while Swin-UNETR achieves 0.610 ± 0.078. The text in the same paragraph states that GynBTNet is 'slightly better than Swin-UNETR (0.610 ± 0.078)', which is numerically false. This inconsistency is load-bearing because the paper's overall superiority claim rests on consistent gains across structures. Please correct the overclaim, either by reporting the sigmoid result as not better on DSC or by adjusting the conclusion to acknowledge the exception.
  2. [Section 3.3, Table 2] The paper claims that self-supervised pretraining 'consistently improved the segmentation performance across all target structures' and 'yielded consistent performance gains across all metrics for most structures.' However, in Table 2 the paired permutation test marks significance († or *) only on the ASD row for all organs; none of the DSC or HD95 improvements are flagged as significant. Since DSC is the primary outcome metric in the paper's own evaluation, the evidence does not establish that pretraining improves segmentation accuracy. The Discussion attributes GynBTNet's superiority over the baselines largely to the pretraining stage; this attribution is not backed by statistically significant DSC or HD95 differences. Please report the actual p-values for all metrics or present the DSC/HD95 improvements as descriptive trends with appropriate caveats.
  3. [Section 3.2, statistical methodology] The one-way ANOVA and Tukey HSD analysis treat nnU-Net, Swin-UNETR, and GynBTNet as independent groups, but the three models are evaluated on the same 29 test cases, creating a paired/repeated-measures structure. Applying independent-sample ANOVA ignores this pairing and can alter the significance conclusions. A repeated-measures ANOVA or paired tests with multiple-comparison correction would be more appropriate. Please clarify why the independent-sample approach was chosen or re-run the analysis with a paired design and state whether the conclusions in Table 1 change.
  4. [Section 2.4 and Section 4] The sparse submanifold convolution is introduced as a key novelty in the self-supervised pretraining stage, but the ablation in Section 3.3 only compares GynBTNet with and without pretraining. No experiment isolates the effect of the sparse submanifold convolution itself, e.g., by comparing it against standard dense convolution under the same masking scheme. Without this ablation, the specific contribution of sparse submanifold convolution cannot be assessed, and the Discussion's claim that it 'enabled the model to learn robust anatomical representations' is unsupported. Please add this comparison or soften the claim.
  5. [Section 2.2.3 and Discussion] The evaluation is performed on 29 test cases from a single institution, with ground-truth contours drawn by one radiation oncologist and reviewed by a senior oncologist. No external validation is provided, and the Discussion acknowledges that stratified analysis by applicator type was not feasible due to the 30-patient sample. Given that the central claim is that GynBTNet improves clinical segmentation accuracy, the generalizability of the results to other institutions, scanners, and contouring practices is a major open question. Please add an explicit limitation paragraph addressing external validity, or provide some form of external or multi-center validation.
minor comments (5)
  1. [Section 2.2.3] The sentence 'Figure 4 shows the detailed distribution of datasets across the three learning stages' appears to reference the wrong figure; Figure 3 shows the dataset distribution, while Figure 4 shows the network architecture. Please correct the cross-reference.
  2. [Throughout] The model name is spelled inconsistently as 'Swin-UNetR' (Table 1) and 'Swin-UNETR' (text and other tables). Please standardize the spelling.
  3. [Section 3.1] The symbol 'HD95%' appears with a percent sign in the metric definition and in Table 1, but the text often uses 'HD95' without the percent sign. Please unify the notation for the 95th percentile Hausdorff distance.
  4. [Section 2.4] The masked patches are described as 'non-overlapping cubic patches of size 7 × 7 × 8'; the three dimensions differ, so 'rectangular cuboids' or 'patches of size 7×7×8' would be more accurate than 'cubic'.
  5. [Figure 1 caption] The caption in the main text and the repeated caption at the end differ in wording; the repeated caption is missing the sentence about circle size. Please ensure a single consistent caption is used.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: GynBTNet's performance claims rest on a held-out 29-case evaluation against external baselines, and the self-supervised pretraining ablation is empirically measured; the only self-citations are non-load-bearing background references.

full rationale

GynBTNet's central claims are empirical: the three-stage training pipeline is evaluated on 29 held-out test cases (Section 2.2.3) against external baselines nnU-Net and Swin-UNETR, and the contribution of self-supervised pretraining is assessed by a direct ablation (GynBTNet without pretraining) using a paired permutation test (Section 3.3). No parameter is fitted to the test labels, no prediction is defined in terms of the outcome, and no uniqueness or forcing theorem is imported from the authors' prior work. The architecture builds on nnU-Net's automatic configuration and the external TotalSegmentator dataset for supervised fine-tuning, so the supervised and task-specific stages are independent of the claimed result. The self-citations [14] and [26] appear only as background references for self-supervised representation learning and for graph neural networks in future work; they are not load-bearing for the segmentation performance or the pretraining-improvement claim. The skeptic concern that Table 2 marks statistical significance only on ASD while the text claims consistent DSC/HD95 improvements is a legitimate evidence-strength issue, but it is not circularity: the ablation numbers are still independently measured on the same held-out test set. Accordingly, no circular step is identified, and the low nonzero score reflects only the presence of minor, non-load-bearing self-citations.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim is empirical and rests on dataset representativeness, ground-truth quality, and transferability of pretraining, plus several hand-selected hyperparameters. No new physical entities are introduced.

free parameters (5)
  • Mask ratio in self-supervised pretraining = 60%
    Chosen in Section 2.4; not optimized or ablated.
  • Mask patch size = 7x7x8
    Non-overlapping cubic patches set in Section 2.4, not justified by experiments.
  • Network scale = 5 stages, 2 blocks per stage, channels 64 to 1024
    Architecture choices in Section 2.3; the 440M parameter scale is a design decision, not derived.
  • Training hyperparameters = lr 1e-4 pretrain, 5e-5 fine-tune; batch sizes 24 and 2; 1000 epochs
    From Section 2.5; hand-selected without ablation.
  • Fine-tuning patch sizes = 112x128x128 and 80x160x160
    Selected for GPU memory and task resolution; may affect results.
assumptions (4)
  • domain assumption Manual contours are accurate enough to serve as ground truth.
    Section 2.2.3 states one radiation oncologist contoured and a senior oncologist reviewed, with MRI as reference. No inter-observer variability or uncertainty is measured.
  • domain assumption The 145-case, 30-patient institutional dataset is representative of GYN-BT cases.
    Section 2.2.3 describes a single institution, one scanner family, and 29 test cases; the paper itself notes that applicator-type stratification was infeasible.
  • domain assumption Self-supervised pretraining on general CT corpora transfers to low-contrast pelvic CT.
    Sections 2.1 and 2.2.1 assume reconstruction on six thousand general CT scans produces useful features for GYN-BT; this is only tested indirectly via ablation.
  • standard math DSC, HD95, and ASD are appropriate and correctly implemented metrics.
    Formulas in Section 3.1 are standard, but no independent verification of the implementation is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Large Convolutional Neural Network for Clinical Target and Multi-organ Segmentation in Gynecologic Brachytherapy with Multi-stage Learning." pith.science (2026). https://pith.science/paper/YSWQ4NXK

@misc{pith2026250601073,
  author       = {Pith},
  title        = {Pith review of: A Large Convolutional Neural Network for Clinical Target and Multi-organ Segmentation in Gynecologic Brachytherapy with Multi-stage Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YSWQ4NXK}},
  note         = {Machine review of arXiv:2506.01073}
}
read the original abstract

Purpose: Accurate segmentation of clinical target volumes (CTV) and organs-at-risk is crucial for optimizing gynecologic brachytherapy (GYN-BT) treatment planning. However, anatomical variability, low soft-tissue contrast in CT imaging, and limited annotated datasets pose significant challenges. This study presents GynBTNet, a novel multi-stage learning framework designed to enhance segmentation performance through self-supervised pretraining and hierarchical fine-tuning strategies. Methods: GynBTNet employs a three-stage training strategy: (1) self-supervised pretraining on large-scale CT datasets using sparse submanifold convolution to capture robust anatomical representations, (2) supervised fine-tuning on a comprehensive multi-organ segmentation dataset to refine feature extraction, and (3) task-specific fine-tuning on a dedicated GYN-BT dataset to optimize segmentation performance for clinical applications. The model was evaluated against state-of-the-art methods using the Dice Similarity Coefficient (DSC), 95th percentile Hausdorff Distance (HD95), and Average Surface Distance (ASD). Results: Our GynBTNet achieved superior segmentation performance, significantly outperforming nnU-Net and Swin-UNETR. Notably, it yielded a DSC of 0.837 +/- 0.068 for CTV, 0.940 +/- 0.052 for the bladder, 0.842 +/- 0.070 for the rectum, and 0.871 +/- 0.047 for the uterus, with reduced HD95 and ASD compared to baseline models. Self-supervised pretraining led to consistent performance improvements, particularly for structures with complex boundaries. However, segmentation of the sigmoid colon remained challenging, likely due to anatomical ambiguities and inter-patient variability. Statistical significance analysis confirmed that GynBTNet's improvements were significant compared to baseline models.

Figures

Figures reproduced from arXiv: 2506.01073 by the authors.

Figure 1
Figure 1. Trends in the number of studies (PubMed) on OAR segmentation, brachytherapy-specific OAR segmentation, and GYN brachytherapy-specific OAR segmentation from 2015 to 2024. The graph highlights the steady growth in general OAR segmentation research, contrasted with the limited growth in brachytherapy-specific studies and the persistently minimal representation of GYN brachytherapy-specific studies. The size of each cir… view at source ↗
Figure 4
Figure 4. Architecture of GynBTNet, featuring a hierarchical encoder-decoder structure with five stages. The network includes a stem block, residual blocks, downsampling blocks, upsampling blocks, and a segmentation head, enabling efficient multi-scale feature extraction and accurate segmentation for HRCTV and OARs [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 6
Figure 6. Sigmoid segmentation results using GynBTNet. The magenta masks represent the ground truth, while the contour lines indicate the segmentation results produced by GynBTNet. Green contours denote cases where the segmentation closely aligns with the ground truth, whereas yellow contours highlight suboptimal segmentations. (d) exhibits over-segmentation, where the predicted region extends beyond the actual sigmoid bounda… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

27 extracted references · 25 canonical work pages

  1. [1]

    Cancer/Radiotherapie, 2024

    Zhu, J., et al., Automatic segmentation of high -risk clinical target volume and organs at risk in brachytherapy of cervical cancer with a convolutional neural network. Cancer/Radiotherapie, 2024. 28(4): p. 354-364

    work page 2024

  2. [2]

    Journal of Applied Clinical Medical Physics, 2020

    Zhang, D., et al., Automatic segmentation and applicator reconstruction for CT‐based brachytherapy of cervical cancer using 3D convolutional neural networks. Journal of Applied Clinical Medical Physics, 2020. 21(10): p. 158-169

    work page 2020

  3. [3]

    Radiation Oncology, 2022

    Li, Z., et al., A deep learning -based self-adapting ensemble method for segmentation in gynecological brachytherapy. Radiation Oncology, 2022. 17(1): p. 152

    work page 2022

  4. [4]

    Journal of Applied Clinical Medical Physics, 2023

    Duprez, D., et al., Self‐configuring nnU‐Net for automatic delineation of the organs at risk and target in high‐dose rate cervical brachytherapy, a low/middle‐income country's experience. Journal of Applied Clinical Medical Physics, 2023. 24(8): p. e13988

    work page 2023

  5. [5]

    Medical physics, 2022

    Cao, Y ., et al., Automatic segmentation of high‐risk clinical target volume for tandem‐and‐ ovoids brachytherapy patients using an asymmetric dual‐path convolutional neural network. Medical physics, 2022. 49(3): p. 1712-1722

    work page 2022

  6. [6]

    Nature Communications, 2024

    Ma, J., et al., Segment anything in medical images. Nature Communications, 2024. 15(1): p. 654

    work page 2024

  7. [7]

    Medclip: Contrastive learning from unpaired medical images and text

    Wang, Z., et al. Medclip: Contrastive learning from unpaired medical images and text . in Proceedings of the Conference on Empirical Methods in Natural Language Processing. Conference on Empirical Methods in Natural Language Processing. 2022

    work page 2022

  8. [8]

    Zhang, S. and D. Metaxas, On the challenges and perspectives of foundation models for medical image analysis. Medical image analysis, 2024. 91: p. 102996

    work page 2024

Show all 27 references
  1. [9]

    Medical Image Analysis, 2022

    Ma, J., et al., Fast and low -GPU-memory abdomen CT organ segmentation: the flare challenge. Medical Image Analysis, 2022. 82: p. 102616

    2022

  2. [10]

    The Cancer Imaging Archive, 2017

    Grossberg, A., et al., Data from head and neck cancer CT atlas. The Cancer Imaging Archive, 2017. 10: p. K9

    2017

  3. [11]

    EBioMedicine, 2020

    Jin, L., et al., Deep-learning-assisted detection and segmentation of rib fractures from CT scans: Development and validation of FracNet. EBioMedicine, 2020. 62

    2020

  4. [12]

    Cancer Imaging Arch, 2015

    Smith, K., et al., Data from CT_COLONOGRAPHY. Cancer Imaging Arch, 2015. 10: p. K9

    2015

  5. [13]

    The Cancer Imaging Archive, 2020

    An, P., et al., Ct images in covid-19 [data set]. The Cancer Imaging Archive, 2020. 10

    2020

  6. [14]

    Radiotherapy and Oncology, 2025

    Li, Y ., et al., Automatic medical imaging segmentation via self -supervising large-scale convolutional neural networks. Radiotherapy and Oncology, 2025. 204: p. 110711

    2025

  7. [15]

    Radiology: Artificial Intelligence, 2023

    Wasserthal, J., et al., Totalsegmentator: Robust segmentation of 104 anatomic structures in ct images. Radiology: Artificial Intelligence, 2023. 5(5)

    2023

  8. [16]

    arXiv preprint arXiv:1809.10486, 2018

    Isensee, F., et al., nnu-net: Self -adapting framework for u -net-based medical image segmentation. arXiv preprint arXiv:1809.10486, 2018

    2018 arXiv

  9. [17]

    arXiv preprint arXiv:2304.06716, 2023

    Huang, Z., et al., Stu-net: Scalable and transferable medical image segmentation models empowered by large-scale supervised pre-training. arXiv preprint arXiv:2304.06716, 2023

    2023 arXiv

  10. [18]

    Tan, M. and Q. Le. Efficientnet: Rethinking model scaling for convolutional neural networks. in International conference on machine learning. 2019. PMLR

    2019

  11. [19]

    Segreg: Segmenting oars by registering mr images and ct annotations

    Zhang, Z., et al. Segreg: Segmenting oars by registering mr images and ct annotations. in 2024 IEEE International Symposium on Biomedical Imaging (ISBI). 2024. IEEE

    2024

  12. [20]

    Medical Physics, 2021

    Huang, S., et al., Integrating multiple MRI sequences for pelvic organs segmentation via the attention mechanism. Medical Physics, 2021. 48(12): p. 7930-7945. 20

    2021

  13. [21]

    Journal of applied clinical medical physics, 2022

    Zabihollahy, F., et al., Fully automated segmentation of clinical target volume in cervical cancer from magnetic resonance imaging with convolutional neural network. Journal of applied clinical medical physics, 2022. 23(9): p. e13725

    2022

  14. [22]

    Medical physics,

    Zabihollahy, F., et al., Fully automated multiorgan segmentation of female pelvic magnetic resonance images with coarse‐to‐fine convolutional neural network. Medical physics,

  15. [23]

    Physics in Medicine & Biology, 2024

    Kim, G., et al., Dual convolution-transformer UNet (DCT -UNet) for organs at risk and clinical target volume segmentation in MRI for cervical cancer brachytherapy. Physics in Medicine & Biology, 2024. 69(21): p. 215014

    2024

  16. [24]

    International Journal of Imaging Systems and Technology, 2022

    Jin, S., et al., Automatic cervical cancer segmentation in multimodal magnetic resonance imaging using an EfficientNet encoder in UNet++ architecture. International Journal of Imaging Systems and Technology, 2022. 33: p. 362 - 377

    2022

  17. [25]

    Medical physics, 2020

    Lei, Y ., et al., CT prostate segmentation based on synthetic MRI‐aided deep attention fully convolution network. Medical physics, 2020. 47(2): p. 530-540

    2020

  18. [26]

    Journal of Medical Imaging, 2024

    Hu, M., et al., MGMT promoter methylation prediction based on multiparametric MRI via vision graph neural network. Journal of Medical Imaging, 2024. 11(1): p. 014503-014503. 21 Figure 1: Trends in the number of studies (PubMed) on OAR segmentation, brachytherapy-specific OAR s...

    2024

  19. [2021]

    7028-7042

    48(11): p. 7028-7042

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.