pith. sign in

arxiv: 2606.10713 · v1 · pith:UOG3CEWEnew · submitted 2026-06-09 · 📡 eess.IV · cs.AI· cs.CV· cs.LG

++nnU-Net: Scaling nnU-Net with Prefix-Based Data Augmentation

Ana Sofia Santos , Andr\'e Ferreira , Gijs Luijten , Naida Solak , Lisle Faray de Paiva , Behrus Hinrichs-Puladi , Jens Kleesiek , Jan Egger
show 1 more author
Victor Alves
This is my paper

Pith reviewed 2026-06-27 11:33 UTC · model grok-4.3

classification 📡 eess.IV cs.AIcs.CVcs.LG
keywords nnU-Netdata augmentationimage registrationmedical image segmentationDice Similarity Coefficient2D datasetswarped images
0
0 comments X

The pith

A registration-based data augmentation step added before nnU-Net training raises Dice scores on 2D medical segmentation tasks.

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

The paper introduces ++nnU-Net to address limited annotated medical data by augmenting datasets through image registration. It applies a two-stage registration to create new warped images and corresponding segmentations prior to the standard nnU-Net pipeline. This approach was tested on five 2D datasets and consistently outperformed the baseline nnU-Net in Dice Similarity Coefficient. The gains reach about 22 percent in the best cases, showing that such augmentation can help in data-scarce medical imaging settings. A sympathetic reader would care because better segmentation with less real data could reduce annotation costs and privacy issues in biomedical applications.

Core claim

The ++nnU-Net framework uses a prefix module that performs two-stage image registration to generate warped images and transformed segmentations, which are then fed into the nnU-Net preprocessing and training. This yields higher Dice Similarity Coefficient scores than the standard nnU-Net across five 2D datasets, with gains up to approximately 22 percent in prominent cases. The method also includes computing disk space, generating binary synthetic masks, and creating checkpoints as part of the scalable pipeline.

What carries the argument

The two-stage registration process that generates warped images and applies transformations to segmentations before preprocessing.

Load-bearing premise

The two-stage registration process generates anatomically feasible warped images and segmentations that improve rather than degrade model training.

What would settle it

Retraining nnU-Net with and without the ++ prefix on the same five 2D datasets and finding no Dice score improvement or a decrease would falsify the performance claim.

Figures

Figures reproduced from arXiv: 2606.10713 by Ana Sofia Santos, Andr\'e Ferreira, Behrus Hinrichs-Puladi, Gijs Luijten, Jan Egger, Jens Kleesiek, Lisle Faray de Paiva, Naida Solak, Victor Alves.

Figure 1
Figure 1. Figure 1: Overview of the augmentation pipeline: (a) data are converted to NIfTI format, renamed, stored in respective [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
read the original abstract

The nnU-Net has demonstrated continuous success in medical segmentation tasks, which heavily rely on the availability and diversity of annotated biomedical data. However, assembling medical imaging cohorts remains challenging due to numerous factors such as privacy regulations and annotation costs. As a result, data augmentation plays a crucial role in increasing data availability while maintaining anatomical feasibility. Hence, we propose the ++nnU-Net, a novel data augmentation module based on image registration that operates prior to preprocessing and training take place. Our framework was evaluated across five different 2D datasets. In this workflow, image data go through a two-stage registration process, generating new warped images. The transformations are then applied to the respective segmentation. In addition, the pipeline computes available disk space, generates supplementary binary synthetic masks and generates checkpoints. We demonstrate that the ++nnU-Net outperforms the nnU-Net baseline, yielding improvements in Dice Similarity Coefficient scores. In the most prominent cases, we observe performance gains of approximately 22\%. These findings highlight the effectiveness of registration-based data augmentation, particularly for 2D medical imaging datasets and suggest that the ++nnU-Net provides a practical and scalable approach for enhancing segmentation performance in data-limited settings. The source code for the ++nnU-Net is available at: https://github.com/sofia-adelie/plusplusnnunet.git

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 3 minor

Summary. The paper introduces ++nnU-Net, a data augmentation module that applies a two-stage image registration process to generate warped images and transformed segmentations prior to nnU-Net preprocessing and training. Evaluated on five 2D medical imaging datasets, it claims consistent outperformance over the nnU-Net baseline with Dice Similarity Coefficient gains reaching approximately 22% in prominent cases, positioning the approach as a scalable solution for data-limited segmentation tasks. The source code is provided via GitHub.

Significance. If the registration-based augmentations are shown to be anatomically valid and the gains are reproducible with proper controls, the work could offer a practical extension to nnU-Net for increasing training diversity without additional annotations. The emphasis on pre-processing augmentation and supplementary mask generation addresses a real constraint in biomedical imaging, though the current presentation supplies insufficient evidence to establish this contribution.

major comments (3)
  1. [Abstract] Abstract: The central claim of ~22% Dice improvement is presented without any description of the registration algorithm (affine vs. deformable, similarity metric, regularization term), post-warp validation (Jacobian checks, label overlap, topology preservation), or comparison against standard nnU-Net augmentations; this directly undermines assessment of whether the reported gains arise from the proposed method rather than uncontrolled factors.
  2. [Abstract] Abstract / Methods (implied): The two-stage registration workflow is asserted to produce 'anatomically feasible' warped images and segmentations, yet no quantitative checks or failure cases are reported; because every downstream Dice number depends on this assumption, the absence of validation constitutes a load-bearing gap for the performance claims.
  3. [Abstract] Abstract: No experimental details are supplied on dataset characteristics, train/validation/test splits, baseline nnU-Net configurations, number of runs, or statistical testing; without these, the cross-dataset superiority claim cannot be evaluated for robustness or confounds such as data leakage.
minor comments (3)
  1. [Abstract] Abstract: Typo in 'prior to preprocessing and training take place' should read 'takes place'.
  2. [Abstract] Abstract: The title refers to 'Prefix-Based Data Augmentation' while the text describes registration-based augmentation; clarify the relationship or intended terminology.
  3. [Abstract] Abstract: The GitHub link is provided but the manuscript does not indicate whether the released code reproduces the exact experimental pipeline and results reported.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments, which identify key gaps in the clarity and supporting evidence of our work. We address each point below and will perform a major revision to incorporate the requested details, validations, and experimental summaries.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim of ~22% Dice improvement is presented without any description of the registration algorithm (affine vs. deformable, similarity metric, regularization term), post-warp validation (Jacobian checks, label overlap, topology preservation), or comparison against standard nnU-Net augmentations; this directly undermines assessment of whether the reported gains arise from the proposed method rather than uncontrolled factors.

    Authors: We agree that the abstract requires additional methodological context to allow proper evaluation. In the revised manuscript we will expand the abstract to include a concise description of the two-stage registration (specifying affine followed by deformable components, the similarity metric, and regularization), note that post-warp validation was performed, and reference the comparison against standard nnU-Net augmentations that appears in the Experiments section. revision: yes

  2. Referee: [Abstract] Abstract / Methods (implied): The two-stage registration workflow is asserted to produce 'anatomically feasible' warped images and segmentations, yet no quantitative checks or failure cases are reported; because every downstream Dice number depends on this assumption, the absence of validation constitutes a load-bearing gap for the performance claims.

    Authors: We acknowledge this as a substantive gap. Although the original submission relied on the design of the registration pipeline to support anatomical feasibility, we will add explicit quantitative validation in the revised version, including Jacobian determinant statistics, label-overlap metrics between warped and original segmentations, and a summary of observed failure cases. revision: yes

  3. Referee: [Abstract] Abstract: No experimental details are supplied on dataset characteristics, train/validation/test splits, baseline nnU-Net configurations, number of runs, or statistical testing; without these, the cross-dataset superiority claim cannot be evaluated for robustness or confounds such as data leakage.

    Authors: We agree that these details must be summarized for the abstract to stand alone. The full manuscript contains a Datasets section and an Experimental Setup section that describe the five 2D datasets, split ratios, nnU-Net baseline configuration, number of runs, and statistical testing. We will insert a compact summary of these elements into the abstract and explicitly address data-leakage safeguards. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical method proposal with direct experimental comparison

full rationale

The paper introduces ++nnU-Net as a registration-based data augmentation pipeline applied before nnU-Net training and reports Dice improvements on five 2D datasets. No equations, parameters fitted to target metrics, or derivation steps appear in the provided text. Performance claims rest on direct empirical comparison to the nnU-Net baseline rather than any reduction to self-referential quantities or self-citations. The two-stage registration assumption is a methodological premise subject to external validation, not a circularity in a claimed derivation chain. This is the standard case of an independent empirical contribution.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract-only review provides no explicit free parameters, axioms, or invented entities; the central claim rests on the unstated domain assumption that registration preserves anatomical validity.

axioms (1)
  • domain assumption Image registration transformations applied to segmentations produce anatomically feasible training examples.
    Invoked implicitly to justify the augmentation step prior to preprocessing.

pith-pipeline@v0.9.1-grok · 5815 in / 1050 out tokens · 20637 ms · 2026-06-27T11:33:30.066621+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

19 extracted references · 3 canonical work pages

  1. [1]

    nnu-net: a self- configuring method for deep learning-based biomedical image segmentation.Nature methods, 18(2):203–211, 2021

    Fabian Isensee, Paul F Jaeger, Simon AA 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

    2021

  2. [2]

    Fully convolutional networks for semantic segmentation

    Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 3431–3440, 2015

    2015

  3. [3]

    U-net: Convolutional networks for biomedical image seg- mentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image seg- mentation. InMedical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234–241. Springer, 2015

    2015

  4. [4]

    The medical segmentation decathlon.Nature Communications, 2022

    Michela Antonelli. The medical segmentation decathlon.Nature Communications, 2022

    2022

  5. [5]

    How much data do you need? an analysis of pelvic multi-organ segmentation in a limited data context.Physical and Engineering Sciences in Medicine, 2025

    Febrio Lunardo. How much data do you need? an analysis of pelvic multi-organ segmentation in a limited data context.Physical and Engineering Sciences in Medicine, 2025

    2025

  6. [6]

    Data augmentation for medical imaging: A systematic literature review.Computers in Biology and Medicine, 2023

    Fabio Garcea. Data augmentation for medical imaging: A systematic literature review.Computers in Biology and Medicine, 2023

    2023

  7. [7]

    Medical image data augmentation: techniques, comparisons and interpretations.Artificial Intelli- gence Review, 56:12561–12605, 2023

    Evgin Goceri. Medical image data augmentation: techniques, comparisons and interpretations.Artificial Intelli- gence Review, 56:12561–12605, 2023

    2023

  8. [8]

    Deep learning using augmentation via registration: 1st place solution to the autoimplant 2020 challenge

    David G Ellis and Michele R Aizenberg. Deep learning using augmentation via registration: 1st place solution to the autoimplant 2020 challenge. InTowards the Automatization of Cranial Implant Design in Cranioplasty: First Challenge, AutoImplant 2020, Held in Conjunction with MICCAI 2020, Lima, Peru, October 8, 2020, Proceedings 1, pages 47–55. Springer, 2...

    2020

  9. [9]

    brainminds.jp/ants-tutorial

    Tutorial: Registering the brain/minds marmoset atlas to new data using ants.https://dataportal. brainminds.jp/ants-tutorial

  10. [10]

    Data augmentation for liver tumor segmentation using structure, texture, and contrast.Bild- verarbeitung für die Medizin 2025, 2025

    Serouj Khajarian. Data augmentation for liver tumor segmentation using structure, texture, and contrast.Bild- verarbeitung für die Medizin 2025, 2025

    2025

  11. [11]

    Gbm-reservoir: Brain tumor (glioblastoma multiforme) mri dataset collection with ground truth segmentation masks author links open overlay panel.Data in Brief, 2025

    Naida Solak. Gbm-reservoir: Brain tumor (glioblastoma multiforme) mri dataset collection with ground truth segmentation masks author links open overlay panel.Data in Brief, 2025

    2025

  12. [12]

    En- hanced data augmentation using synthetic data for brain tumour segmentation, 2024

    André Ferreira, Naida Solak, Jianning Li, Philipp Dammann, Jens Kleesiek, Victor Alves, and Jan Egger. En- hanced data augmentation using synthetic data for brain tumour segmentation, 2024

    2024

  13. [13]

    100+ 2d us images and tumor segmentation masks.https://doi.org/10.13140/RG.2.2.36586

    Jan Egger. 100+ 2d us images and tumor segmentation masks.https://doi.org/10.13140/RG.2.2.36586. 77761

    work page doi:10.13140/rg.2.2.36586

  14. [14]

    Dataset of breast ultrasound images.Data in Brief, 2020

    Walid Al-Dhabyani. Dataset of breast ultrasound images.Data in Brief, 2020

    2020

  15. [15]

    Killeen, M

    B. Killeen, M. Liu, P.-C. Ku, S. Yudi, Y . Liu, S. Yibulayimu, G. Zhu, X. Wu, C. Zhao, Y . Wang, M. Armand, and M. Unberath. Pengwin task 2: Pelvic fragment segmentation on synthetic x-ray images.https://doi.org/ 10.5281/zenodo.10913196

    work page doi:10.5281/zenodo.10913196

  16. [16]

    Breast ultrasound images dataset (busi).https://scholar.cu.edu.eg/?q=afahmy/ pages/dataset

    Walid Al-Dhabyani. Breast ultrasound images dataset (busi).https://scholar.cu.edu.eg/?q=afahmy/ pages/dataset

  17. [17]

    Arcade: Automatic region-based coronary artery disease diagnostics using x-ray angiography images dataset.https://zenodo.org/records/10390295

    Maxim popov. Arcade: Automatic region-based coronary artery disease diagnostics using x-ray angiography images dataset.https://zenodo.org/records/10390295

    work page arXiv

  18. [18]

    Teeth segmentation on dental x-ray images, 2023

    Humans In The Loop. Teeth segmentation on dental x-ray images, 2023

    2023

  19. [19]

    Popov, A

    M. Popov, A. Amanturdieva, and N. Zhaksylyk. Dataset for automatic region-based coronary artery disease diagnostics using x-ray angiography images.Scientific Data, 2024. 7

    2024