Pith. sign in

REVIEW 3 major objections 5 minor 16 references

Comparative Analysis of nnUNet and MedNeXt for Head and Neck Tumor Segmentation in MRI-guided Radiotherapy

T0 review · 3 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read Single MedNeXt model wins head-and-neck MRI tumor segmentation task

desk verdict First published HNTS-MRG24 challenge report with an externally supported 1st-place ranking and a genuinely useful but registration-dependent mask-ablation finding. read the letter →

arxiv 2411.14752 v1 pith:ZSPJEYMN submitted 2024-11-22 eess.IV cs.CVcs.LG

Nikoo Moradi , André Ferreira , Behrus Puladi , Jens Kleesiek , Emad Fatemizadeh , Gijs Luijten , Victor Alves , Jan Egger This is my paper · ORCID
classification eess.IVcs.CVcs.LG
keywords HNTS-MRG24MICCAI24nnUNetMedNeXtheadandneckcancerMRI-guidedradiotherapytumorsegmentationaggregatedDicesimilarity
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

This paper reports the winning solution to the head-and-neck MRI tumor segmentation challenge's first task (pre-radiotherapy scans) and an eighth-place solution to its second task (mid-radiotherapy scans). The authors claim that a small MedNeXt model with 3×3×3 kernels outperformed every nnUNet configuration on the pre-treatment task, reaching an aggregated Dice Similarity Coefficient of 0.8254 on the final test set. For the mid-treatment task, they claim that adding the registered pre-treatment image and its segmentation mask as extra input channels clearly improves both models compared to using mid-treatment images alone. This matters because MRI-guided adaptive radiotherapy needs fast, accurate tumor delineation at every treatment session, and the result shows that prior time-point masks carry useful localization information that can be supplied to the network directly.

What carries the argument

The mechanism that carries the mid-RT argument is input-channel composition: the network receives the mid-RT T2-weighted scan together with the registered pre-RT scan and the registered pre-RT segmentation mask (datasets 506 and 507), effectively using the earlier time point's expert annotation as a spatial prior for where the tumor can be found at mid-treatment. For Task 1, the decisive model is MedNeXt Small with kernel size 3, a convolutional architecture with 32 channels, an expansion ratio of 2, and 2 blocks per stage, whose standalone predictions outperformed every nnUNet combination. The ranking itself is carried by the aggregated Dice Similarity Coefficient ($DSC_{agg}$), which sums volumes across the whole test set instead of averaging per-sample scores, as defined in Equation 1 of the paper.

What would settle it

Train the Task 2 pipelines on a cohort with independently verified high-accuracy registration, then repeat the same training with artificially perturbed registration; if the benefit of adding pre-RT masks persists under misalignment, the conclusion is not registration-dependent. A simpler check is to re-run the dataset 504 versus 506 comparison on a test set whose pre-RT-to-mid-RT registration was recomputed with a different algorithm and see whether the mask-channel advantage remains.

Watch

Extended reading notes

Core claim

The central claim, stated on the paper's own terms, is that on the challenge's 150-patient dataset a single MedNeXt Small (kernel 3) model produces the best pre-RT segmentation, beating all nnUNet ensembles and achieving an aggregated Dice of 0.8254 to rank first in Task 1, while an nnUNet ensemble of Full Resolution and Cascade models ranks eighth in Task 2 with 0.7005. The authors also establish that for mid-RT segmentation, including the registered pre-RT image and registered pre-RT mask as extra input channels (their dataset 506) gives the highest aggregated Dice for nnUNet, and that including the pre-RT mask is the key factor: inputs with masks outperform mid-RT-only inputs, while adding the registered image alone does not help. They further report that averaging nnUNet and MedNeXt probability maps lowers performance, and that the larger MedNeXt variants and kernel-5 versions collapsed during training, so the comparison is limited to the small kernel-3 MedNeXt.

Load-bearing premise

The Task 2 result assumes that the challenge-provided registration aligning the pre-RT images to the mid-RT images is accurate enough that the registered masks and images, when added as input channels, contribute real anatomical correspondence rather than misalignment noise.

Editorial extensions

If this is right

  • If the Task 1 result holds, a single MedNeXt Small kernel-3 model is sufficient to exceed multi-model nnUNet ensembles on pre-RT head-and-neck scans, which could simplify deployment.
  • If the Task 2 result holds, adaptive radiotherapy workflows can store the planning-time contours and feed them as input channels at each fraction, improving mid-treatment segmentation without architectural changes.
  • The observation that adding the registered pre-RT image alone does not help, while adding the mask does, implies that the meaningful prior is the annotated location, not the image appearance.
  • The drop in performance from averaging nnUNet and MedNeXt probability maps indicates that naive cross-architecture ensembles are unreliable for this task, so future ensembles should align or weight the predictions carefully.
  • The collapse of larger MedNeXt variants and kernel-5 models on this dataset shows that those configurations need stabilized training procedures or more data before they can be used for small, unbalanced medical segmentation sets.

Reading between the lines

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

  • One untested extension is per-label model selection: the per-label Dice tables suggest MedNeXt may be stronger on GTVp while nnUNet is stronger on GTVn, so assigning each label its best model could exceed either ensemble.
  • If the mask-channel benefit generalizes, longitudinal segmentation could be framed as conditioning on a prior contour, suggesting connections to interactive or prompt-based segmentation where the previous mask acts as a prompt.
  • The poor transfer from the brain MRI pretraining data suggests cross-anatomy pretraining for segmentation needs explicit domain adaptation; testing progressive unfreezing or contrastive alignment between T1w brain and T2w head-and-neck scans would be a concrete next experiment.
  • The reported training instability of larger MedNeXt variants on a 115-sample filtered dataset hints that hyperparameter or optimizer adjustments, rather than architecture capability, may unlock those models on small medical datasets.
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

3 major / 5 minor

Summary. This paper reports the HNTS-MRG24 MICCAI challenge solution from team TUMOR. For Task 1 (pre-RT segmentation of GTVp and GTVn), the authors pretrain nnUNet and MedNeXt variants on mid-RT and registered pre-RT images, then fine-tune on original pre-RT images; the final Task 1 submission, MedNeXt Small with kernel size 3, achieved 1st place with DSCagg 0.8254 on the 50-case test set. For Task 2 (mid-RT segmentation), the authors evaluate multi-channel inputs that combine mid-RT, registered pre-RT images, and registered pre-RT segmentation masks; the final submission, an nnUNet FullRes+Cascade ensemble trained on Dataset 506, ranked 8th with DSCagg 0.7005. The paper also compares model families and ablates the contribution of the registered pre-RT masks.

Significance. If the leaderboard claims are correct, the paper documents a top-performing, publicly available pipeline for a clinically relevant segmentation task, and the challenge ranking provides external validation. The more interesting scientific content is the Task 2 ablation showing that registered pre-RT masks improve mid-RT segmentation. That conclusion, however, is conditional on the quality of the challenge-provided registration and is partly confounded by incomplete MedNeXt training. The paper is honest about training collapses and the invalidity of the pretraining comparison, which is a strength, but the central secondary claim needs additional support before it can be considered transferable.

major comments (3)
  1. [Section 2.4, Table 6, Section 4] The central Task 2 conclusion that adding registered pre-RT segmentation masks improves mid-RT segmentation is only cleanly supported for nnUNet. The MedNeXt rows in Table 6 are confounded: the Dataset 506 result is produced by models that did not complete the planned 1000 epochs on folds 0, 3, and 4 (Table 5), and the Dataset 516 result comes from a filtered 115-case subset that removes cases with zero ground truth. The sentence in Section 4 that 'Both nnUNet and MedNeXt models trained on datasets which included segmentation masks ... performed better' therefore mixes a clean nnUNet ablation with a partially trained and differently distributed MedNeXt comparison. Please either retrain MedNeXt on the full Dataset 506 or separate the nnUNet ablation from the exploratory MedNeXt comparison and label the latter as such.
  2. [Section 2.4, Table 6, Section 4] The paper's secondary scientific claim that registered pre-RT masks 'significantly improve' mid-RT segmentation is conditional on the spatial accuracy of the challenge-provided registration. No registration quality metric is reported, and no ablation controls for misalignment, such as feeding the unregistered pre-RT mask or a randomly perturbed mask as a negative control. Without such evidence, the observed gain may reflect the model copying or refining an input that is informative only because the registration happens to align pre-RT and mid-RT anatomy; the conclusion would not transfer to other registration pipelines. Please add a registration quality measure (e.g., landmark or overlap statistics) or a misalignment ablation.
  3. [Section 3 (Task 1) and Section 4] The title promises a comparative analysis of nnUNet and MedNeXt, but only the MedNeXt Small (Kernel 3) variant trained successfully; Small (Kernel 5), Large (Kernel 3), and Large (Kernel 5) all collapsed in both tasks. The claim in Section 4 that 'MedNeXt proved to be a stronger candidate for reliable segmentation compared to nnUNet' is therefore supported only for this single variant on Task 1. Please narrow the conclusion to 'MedNeXt Small (Kernel 3)' and explicitly state that the comparison is between one MedNeXt configuration and the complete nnUNet family, as the paper itself acknowledges later in the same section.
minor comments (5)
  1. [Abstract and throughout] There are several typos, including 'therfore', 'T ask', 'of0.8254', and 'ResEnd' instead of 'ResEnc'; these should be corrected.
  2. [Section 2.1] Please clarify the data split: the text says the dataset comprises 150 MRI scans and later refers to '50 test patients'; the number of training and validation cases used for Tables 1-8 is not stated.
  3. [Section 3, Tables 1 and 6] DSCagg is reported as a single aggregate with no uncertainty; adding bootstrap confidence intervals or paired per-case tests would make the model-comparison statements more robust.
  4. [Section 2.4] The statement that 'every possible combination of nnUNet models was aggregated' is not fully reflected in Table 1, which lists only some combinations; either list all combinations or revise the wording.
  5. [Section 2.5, Eqs. (1)-(2)] The notation for DSCagg and per-case DSC is easy to confuse; please clarify that Eq. (1) aggregates over the full set while Eq. (2) is per image i, and define the index ranges explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper is an empirical challenge report whose ranking claims are externally measured and whose ablations are direct dataset comparisons, not derivations from fitted parameters or self-citations.

full rationale

This paper is an empirical challenge report, not a derivation. The central claims (Task 1 first place with DSCagg 0.8254; Task 2 eighth with 0.7005) are benchmark results measured on the HNTS-MRG24 test set and are externally grounded in the challenge leaderboard; they are not the output of a fitted parameter that is later relabeled as a prediction. The Task 2 ablation in Section 2.4 and Table 6 compares four input-channel configurations (Datasets 504-507) under the same training protocol. The conclusion that adding registered pre-RT segmentation masks improves mid-RT DSCagg is a direct empirical comparison: no equation in the paper equates the added input channels with the metric being predicted, and the masks are ground-truth contours rather than predictions derived from the model. Using the validation set to select the best dataset and model is standard challenge practice and does not make the reported test-phase scores circular. References to nnUNet and MedNeXt are external architecture citations; the present authors are not among the cited MedNeXt/nnUNet author lists, so there is no load-bearing self-citation. The paper's reliance on challenge-provided registration is a generalizability limitation, not circularity, because the registered inputs are not constructed from the mid-RT outcome labels. No load-bearing step reduces by construction to its own inputs, so the appropriate finding is no significant circularity.

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

No new entities or forces are introduced. The analysis leans on standard deep learning toolkit assumptions and challenge-provided data alignment; these axioms are reasonable but untested in the paper.

assumptions (4)
  • domain assumption Registered pre-RT images are spatially aligned to mid-RT images accurately enough for multi-channel input.
    Used in Task 2 dataset construction (Section 2.4) and in the conclusion that pre-RT masks help mid-RT segmentation.
  • domain assumption Ground truth segmentation masks from the challenge are correct and consistent across labels.
    All DSC scores are computed against these masks; no inter-observer or quality check is reported (Section 2.1).
  • domain assumption DSCagg (Eq. 1) is an appropriate measure of segmentation quality for ranking.
    The paper adopts the challenge metric without questioning its clinical or statistical properties (Section 2.5).
  • domain assumption Default nnUNet planner and MedNeXt architecture settings are suitable for this dataset.
    The paper relies on framework defaults and prior MedNeXt design choices (Section 2.2).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Comparative Analysis of nnUNet and MedNeXt for Head and Neck Tumor Segmentation in MRI-guided Radiotherapy." pith.science (2026). https://pith.science/paper/ZSPJEYMN

@misc{pith2026241114752,
  author       = {Pith},
  title        = {Pith review of: Comparative Analysis of nnUNet and MedNeXt for Head and Neck Tumor Segmentation in MRI-guided Radiotherapy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZSPJEYMN}},
  note         = {Machine review of arXiv:2411.14752}
}
read the original abstract

Radiation therapy (RT) is essential in treating head and neck cancer (HNC), with magnetic resonance imaging(MRI)-guided RT offering superior soft tissue contrast and functional imaging. However, manual tumor segmentation is time-consuming and complex, and therfore remains a challenge. In this study, we present our solution as team TUMOR to the HNTS-MRG24 MICCAI Challenge which is focused on automated segmentation of primary gross tumor volumes (GTVp) and metastatic lymph node gross tumor volume (GTVn) in pre-RT and mid-RT MRI images. We utilized the HNTS-MRG2024 dataset, which consists of 150 MRI scans from patients diagnosed with HNC, including original and registered pre-RT and mid-RT T2-weighted images with corresponding segmentation masks for GTVp and GTVn. We employed two state-of-the-art models in deep learning, nnUNet and MedNeXt. For Task 1, we pretrained models on pre-RT registered and mid-RT images, followed by fine-tuning on original pre-RT images. For Task 2, we combined registered pre-RT images, registered pre-RT segmentation masks, and mid-RT data as a multi-channel input for training. Our solution for Task 1 achieved 1st place in the final test phase with an aggregated Dice Similarity Coefficient of 0.8254, and our solution for Task 2 ranked 8th with a score of 0.7005. The proposed solution is publicly available at Github Repository.

Figures

Figures reproduced from arXiv: 2411.14752 by the authors.

Figure 1
Figure 1. A sample pre-RT image (Case 78) with its corresponding segmentation. The green label represents GTVp (label=1), while the yellow label represents GTVn (label=2). The images show axial, coronal, and sagittal views, along with a 3D rendering of the segmented tumors. We used all 500 cases of the BraTS dataset specifically for Task 1 of the challenge and applied it as pretraining data for our models. 2.2 Networks The ef… view at source ↗
Figure 2
Figure 2. Comparison of predicted segmentations of two pre-RT samples (Case 78 and 166) for Task 1. The left image shows the prediction from MedNeXt with the best DSCagg, the middle image shows the prediction from the average ensemble of nnUNet and MedNeXt, which had the lowest DSCagg, and the right image shows the ground truth segmentation. The green label represents GTVp (label=1), and the yellow label represents GTVn (labe… view at source ↗
Figure 3
Figure 3. Comparison of segmentation predictions of sample mid-RT (Case 78 and 166) for Task 2. The left image shows the prediction from ensemble of nnUNet Cascade and FullRes with the best DSCagg, the middle image shows the prediction from the average ensemble of nnUNet and MedNeXt, which had the lowest DSCagg, and the right image shows the ground truth segmentation. The green label represents GTVp (label=1), and the yellow … view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

16 extracted references · 13 canonical work pages

  1. [1]

    Kiser, Benjamin D

    Kendall J. Kiser, Benjamin D. Smith, Jihong Wang, and Clifton D. Fuller. Après mois, le déluge: Preparing for the coming data flood in the mri-guided radiotherapy era.Frontiers in Oncology, 9:983, 2019

    work page 2019

  2. [2]

    Pollard, Zhifei Wen, Ramaswamy Sadagopan, Jihong Wang, and Geoffrey S

    Julianne M. Pollard, Zhifei Wen, Ramaswamy Sadagopan, Jihong Wang, and Geoffrey S. Ib- bott. The future of image-guided radiotherapy will be mr guided.British Journal of Radiology, 90(1073):20160667, 2017

    work page 2017

  3. [3]

    Overview of the hecktor challenge at miccai 2022: Automatic head and neck tumor segmenta- tion and outcome prediction in pet/ct

    Vincent Andrearczyk, Valentin Oreiller, Moamen Abobakr, Azadeh Akhavanallaf, et al. Overview of the hecktor challenge at miccai 2022: Automatic head and neck tumor segmenta- tion and outcome prediction in pet/ct. InHead and Neck Tumor Segmentation and Outcome Prediction. HECKTOR 2022. LNCS, volume 13626, pages 1–30. Springer, Cham, 2023

    work page 2022

  4. [4]

    Segrap2023: A benchmark of organs-at-risk and gross tumor volume segmentation for radiotherapy planning of nasopharyngeal carcinoma

    Xiangde Luo, Jia Fu, Yunxin Zhong, Shuolin Liu, Bing Han, Mehdi Astaraki, Simone Bendaz- zoli, Iuliana Toma-Dasu, et al. Segrap2023: A benchmark of organs-at-risk and gross tumor volume segmentation for radiotherapy planning of nasopharyngeal carcinoma. In MICCAI SegRap 2023, 2023

    work page 2023

  5. [5]

    Jaeger, Simon A

    Fabian Isensee, Paul F. Jaeger, Simon A. Kohl, Jens Petersen, and Klaus H. Maier-Hein. nnu- net: a self-configuring method for deep learning-based biomedical image segmentation.Nature methods, 18(2):203–211, 2021

    work page 2021

  6. [6]

    Jaeger, and Klaus H

    Subhrajit Roy, Georg Koehler, Christian Ulrich, Michael Baumgartner, Jens Petersen, Fabian Isensee, Paul F. Jaeger, and Klaus H. Maier-Hein. Mednext: Transformer-driven scaling of convnets for medical image segmentation. InInternational Conference on Medical Image Com- puting and Computer-Assisted Intervention (MICCAI), 2023

    work page 2023

  7. [7]

    Jaeger, Simon A

    Fabian Isensee, Paul F. Jaeger, Simon A. Kohl, Jens Petersen, and Klaus H. Maier-Hein. nnu- net: a self-configuring method for deep learning-based biomedical image segmentation.Nature Methods, pages 1–9, 2020

    work page 2020

  8. [8]

    A generic ensemble based deep convolutional neural network for semi-supervised medical image segmentation

    Ruizhe Li, Dorothee Auer, Christian Wagner, and Xin Chen. A generic ensemble based deep convolutional neural network for semi-supervised medical image segmentation.arXiv preprint arXiv:2004.07995, 2020

    work page Pith review arXiv 2004

Show all 16 references
  1. [9]

    Emre Celebi, Stephen Dusza, David Gutman, Brian Helba, Aadi Kalloo, Konstantinos Liopyris, and Michael Marchetti

    Noel Codella, Veronica Rotemberg, Philipp Tschandl, M. Emre Celebi, Stephen Dusza, David Gutman, Brian Helba, Aadi Kalloo, Konstantinos Liopyris, and Michael Marchetti. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging co...

    2018 arXiv

  2. [10]

    Theham10000dataset,alargecollection of multi-source dermatoscopic images of common pigmented skin lesions

    PhilippTschandl,CliffRosendahl,andHaraldKittler. Theham10000dataset,alargecollection of multi-source dermatoscopic images of common pigmented skin lesions. Scientific Data, 5:180161, 2018

    2018

  3. [11]

    Fully automatic segmentation of gross target volume and organs-at-risk for radiotherapy planning of nasopharyngeal carcinoma

    Mehdi Astaraki, Simone Bendazzoli, and Iuliana Toma-Dasu. Fully automatic segmentation of gross target volume and organs-at-risk for radiotherapy planning of nasopharyngeal carcinoma. arXiv:2310.02972, 2023

    2023 arXiv

  4. [12]

    Automated head and neck tumor segmentation from 3d pet/ct: Hecktor 2022 challenge report

    Andriy Myronenko, Md Mahfuzur Rahman Siddiquee, Dong Yang, Yufan He, and Daguang Xu. Automated head and neck tumor segmentation from 3d pet/ct: Hecktor 2022 challenge report. arXiv preprint arXiv:2209.10809, 2022. Comparison of Automatic Segmentation Frameworks 15

    2022 arXiv

  5. [13]

    3d slicer

    Steve Pieper, Michael Halle, and Ron Kikinis. 3d slicer. In 2004 2nd IEEE International Symposium on Biomedical Imaging: Nano to Macro (IEEE Cat No. 04EX821), pages 632–635 Vol. 1, 2004

    2004

  6. [14]

    Brain tumor segmentation (brats) challenge 2024: Meningioma radiotherapy planning automated segmentation.arXiv preprint arXiv:2405.18383, 2024

    Dominic LaBella, Katherine Schumacher, Michael Mix, Kevin Leu, Shan McBurney-Lin, Pierre Nedelec, et al. Brain tumor segmentation (brats) challenge 2024: Meningioma radiotherapy planning automated segmentation.arXiv preprint arXiv:2405.18383, 2024

    2024 arXiv

  7. [15]

    Prior, and Adrien Depeursinge

    Vincent Andrearczyk, Valentin Oreiller, Mario Jreige, Joël Castelli, John O. Prior, and Adrien Depeursinge. Segmentation and classification of head and neck nodal metastases and primary tumors in pet/ct. In2022 44th Annual International Conference of the IEEE Engineering in Me...

    2022

  8. [16]

    Lee R. Dice. Measures of the amount of ecologic association between species. Ecology, 26(3):297–302, 1945

    1945

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.