arxiv: 2605.01075 · v1 · submitted 2026-05-01 · 💻 cs.CV

Recognition: unknown

Neighbor2Inverse: Self-Supervised Denoising for Low-Dose Region-of-Interest Phase Contrast CT

Daniela Pfeiffer, Daniel Frey, Florian Schaff, Franz Pfeiffer, Jannis Ahlers, Johannes B. Thalhammer, Kaye S. Morgan, Lorenzo D'Amico, Lucy Costello, Marcus Kitchen, Martin Donnelley, Ronan Smith, Sebastian Peterhansl, Tina Dorosti

Authors on Pith no claims yet

Pith reviewed 2026-05-09 19:13 UTC · model grok-4.3

classification 💻 cs.CV

keywords self-supervised denoisinglow-dose CTphase-contrast imagingpropagation-based imagingregion-of-interest CTNeighbor2Neighbor

0 comments

The pith

Neighbor2Inverse enables self-supervised denoising for low-dose region-of-interest phase contrast CT by training on subsampled projection pairs.

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

The paper introduces Neighbor2Inverse, a self-supervised denoising framework for low-dose propagation-based X-ray phase-contrast CT that avoids the need for paired low- and high-dose datasets. Each noisy projection is subsampled into two variants that preserve structure but have independent noise, then reconstructed separately to create training pairs for a denoising network in the image domain. In ROI PBI CT experiments, it demonstrates superior noise suppression and detail preservation over analytical and self-supervised baselines, measured by improved contrast-to-noise ratio, spatial resolution, and quality metrics, with competitive results on clinical CT under low-dose simulation. This approach supports safer medical imaging by enabling effective denoising without high-dose references.

Core claim

Neighbor2Inverse achieves superior noise suppression while preserving fine structural details in low-dose region-of-interest phase contrast CT experiments, as shown by better contrast-to-noise ratio, spatial resolution, and composite image quality metrics, through self-supervised training on image pairs from separately reconstructed subsampled projections.

What carries the argument

The Neighbor2Inverse method, which builds on Neighbor2Neighbor by subsampling noisy projections into two independent-noise variants for separate reconstruction and image-domain self-supervised denoising training.

If this is right

Superior performance in noise suppression and detail preservation for PBI CT.
Competitive denoising on simulated low-dose clinical CT data.
Avoids requirement for paired high-dose training data.
Generalizes the Neighbor2Neighbor principle to the PBI-CT inverse problem.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The method could enable lower radiation doses in clinical phase-contrast lung imaging.
It may inspire similar self-supervised approaches for other CT modalities or noise reduction tasks.
Validation on actual clinical low-dose scans, rather than simulations, would be a natural next test.

Load-bearing premise

Subsampling each noisy projection produces two variants with independent noise realizations but identical structural information.

What would settle it

If applying the method results in no improvement in contrast-to-noise ratio or loss of fine details in the denoised images relative to baselines, the claim would be falsified.

Figures

Figures reproduced from arXiv: 2605.01075 by Daniela Pfeiffer, Daniel Frey, Florian Schaff, Franz Pfeiffer, Jannis Ahlers, Johannes B. Thalhammer, Kaye S. Morgan, Lorenzo D'Amico, Lucy Costello, Marcus Kitchen, Martin Donnelley, Ronan Smith, Sebastian Peterhansl, Tina Dorosti.

**Figure 1.** Figure 1: A The Neighbor subsampling algorithm. B The Neighbor2Inverse pipeline C with a data fidelity regularization instead of the regularization based on different underlying signals. 2.3.2 Neighbor2Inverse The proposed Neighbor2Inverse method (Fig. 1B) extends Neighbor2Neighbor to the tomographic inverse problem by training the denoising network directly in the image domain, where spatial correlations can be mor… view at source ↗

**Figure 2.** Figure 2: Comparison of denoising methods applied to two 15 ms reconstructed propagation-based CT images from the view at source ↗

**Figure 3.** Figure 3: Comparison of denoising methods applied to clinical CT data with simulated noise added at projection level. view at source ↗

**Figure 4.** Figure 4: Denoising results of various Neighbor2Inverse approaches applied to two 15 ms reconstructed propagation view at source ↗

**Figure 5.** Figure 5: Sensitivity Analysis. The initial heuristically chosen weight view at source ↗

**Figure 6.** Figure 6: Denoising results of various Neighbor2Inverse and the Neighbor2Neighbor approaches applied to a sparse 15 view at source ↗

**Figure 7.** Figure 7: Denoising results of Neighbor2Inverse with varying exposure times and projection views. The displayed images view at source ↗

**Figure 8.** Figure 8: Quantitative evaluation of the different denoising methods applied on view at source ↗

read the original abstract

Propagation-based X-ray phase-contrast imaging (PBI) enables high-contrast visualization of lung structures and holds strong medical potential. However, safe translation to the clinic will require a substantial radiation dose reduction, which inevitably increases image noise. Supervised convolutional-neural-network-based denoising can restore image quality but depends on paired low- and high-dose datasets, which are rarely available in practice. Self-supervised methods avoid this limitation, yet most are not well adapted to the inverse problem of PBI computed tomography (CT). We introduce Neighbor2Inverse, a self-supervised denoising framework designed for low-dose PBI-CT that generalizes to clinical CT. Building on the Neighbor2Neighbor principle, each noisy projection is subsampled into two variants that preserve structural information but contain independent noise realizations. These are reconstructed separately, and the resulting pairs are used to train a denoising network directly in the image domain. We benchmark the proposed method against established analytical and self-supervised denoising approaches. In region-of-interest PBI CT experiments, Neighbor2Inverse achieves superior noise suppression while preserving fine structural details, as demonstrated by improved contrast-to-noise ratio, spatial resolution, and composite image quality metrics. Competitive performance is also observed on clinical CT data under simulated low-dose conditions. This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible. Code, data, and interactive figures are available at https://github.com/J-3TO/Neighbor2Inverse.

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.

Desk Editor's Note private letter to a colleague

Neighbor2Inverse adapts Neighbor2Neighbor to PBI-CT via projection subsampling and image-domain training, but the noise-independence assumption after nonlinear reconstruction looks untested and potentially fragile.

read the letter

The paper takes the Neighbor2Neighbor self-supervised idea and applies it to low-dose region-of-interest propagation-based phase-contrast CT. Each noisy projection gets split into two variants that keep the same underlying structure but should carry independent noise; those are reconstructed separately and the resulting image pairs train a denoising network directly in image space. They also run the same pipeline on simulated low-dose clinical CT data for comparison. This avoids the paired clean-noisy datasets that supervised methods need, which is a real practical advantage for phase-contrast work where high-dose references are hard to get. The reported metrics show better contrast-to-noise ratio, spatial resolution, and overall quality scores than the analytical and self-supervised baselines they tested against in the ROI PBI-CT experiments, and competitive numbers on the clinical CT side. Public code and data are a plus for anyone wanting to try it out. The soft spot is the central assumption that subsampling projections produces pairs with identical structure and uncorrelated noise once phase retrieval and reconstruction are done. Phase retrieval is nonlinear, ROI truncation often requires interpolation or padding, and nothing in the abstract or method description shows a direct check on post-reconstruction noise correlation or structural match between the pairs. If those conditions slip, the training signal becomes unreliable and the claimed gains could be overstated. A reader working on dose reduction for phase-contrast CT or on self-supervised inverse-problem denoising would get value from the benchmarks and the open implementation. The paper is coherent on its own terms and engages the literature without obvious internal contradictions, so it deserves a serious referee to pressure-test the noise-independence claim and the experimental controls. I would send it to peer review.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes Neighbor2Inverse, a self-supervised denoising framework for low-dose region-of-interest propagation-based phase-contrast CT (PBI-CT) that generalizes to clinical CT. It adapts the Neighbor2Neighbor principle by subsampling each noisy projection into two variants that preserve structure but contain independent noise, reconstructing the pairs separately, and training a denoising network directly in the image domain. Experiments on ROI PBI-CT data report superior noise suppression and detail preservation via improved contrast-to-noise ratio, spatial resolution, and composite quality metrics compared to analytical and self-supervised baselines, with competitive performance under simulated low-dose clinical CT conditions. Code and data are made available.

Significance. If the core training-pair assumption holds, the method could enable effective denoising for PBI-CT without requiring scarce paired low/high-dose data, supporting safer clinical translation through dose reduction while preserving fine lung structures. The explicit generalization test on clinical CT and the public release of code, data, and interactive figures are strengths that aid reproducibility and adoption.

major comments (2)

[§3 (Method)] §3 (Method), pair-generation step: the claim that subsampling each projection produces pairs with identical underlying structure but uncorrelated noise after separate phase retrieval and reconstruction is central to the self-supervised loss, yet no direct validation (e.g., measured noise correlation coefficient or structural similarity between the reconstructed pair images) is provided. Nonlinear phase retrieval (Paganin-type or equivalent) and ROI truncation/interpolation can introduce correlations or structural mismatches, rendering the loss unreliable; this must be quantified before the superiority claims can be accepted.
[§4 (Experiments)] §4 (Experiments), quantitative results: the reported gains in CNR, spatial resolution, and composite metrics on ROI PBI-CT lack accompanying statistical analysis (number of independent volumes, p-values, or standard deviations across runs), and the precise baseline implementations (including which self-supervised methods and their hyperparameters) are insufficiently specified to allow reproduction or assessment of whether the improvements are robust.

minor comments (2)

[Abstract and §4] The abstract and method sections refer to 'composite image quality metrics' without an explicit definition or formula in the main text; a short clarification or reference would improve readability.
[Figures in §4] Figure captions and axis labels in the experimental results could more clearly distinguish between the proposed method and each baseline to aid quick visual comparison.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed review. We address each major comment below and will revise the manuscript to incorporate the suggested improvements for greater rigor and reproducibility.

read point-by-point responses

Referee: [§3 (Method)] §3 (Method), pair-generation step: the claim that subsampling each projection produces pairs with identical underlying structure but uncorrelated noise after separate phase retrieval and reconstruction is central to the self-supervised loss, yet no direct validation (e.g., measured noise correlation coefficient or structural similarity between the reconstructed pair images) is provided. Nonlinear phase retrieval (Paganin-type or equivalent) and ROI truncation/interpolation can introduce correlations or structural mismatches, rendering the loss unreliable; this must be quantified before the superiority claims can be accepted.

Authors: We agree that direct empirical validation of the core assumption is necessary to support the self-supervised framework. While the projection subsampling is designed to occur prior to phase retrieval and reconstruction to maintain independent noise while preserving structure, we acknowledge that nonlinear phase retrieval and ROI processing steps could introduce unintended correlations. In the revised manuscript, we will add quantitative validation in Section 3, including measured noise correlation coefficients and structural similarity (SSIM) values between the reconstructed image pairs, to confirm that the pairs remain suitable for the loss function. revision: yes
Referee: [§4 (Experiments)] §4 (Experiments), quantitative results: the reported gains in CNR, spatial resolution, and composite metrics on ROI PBI-CT lack accompanying statistical analysis (number of independent volumes, p-values, or standard deviations across runs), and the precise baseline implementations (including which self-supervised methods and their hyperparameters) are insufficiently specified to allow reproduction or assessment of whether the improvements are robust.

Authors: We acknowledge that the current presentation of results would benefit from additional statistical details and explicit baseline specifications to strengthen the claims and support reproducibility. In the revised version, we will report standard deviations across independent volumes or multiple runs for all metrics, specify the number of test volumes used, and include p-values for the reported improvements. We will also expand the experimental section with precise descriptions of all baseline methods, including their implementations, hyperparameters, and any adaptations for the PBI-CT data. revision: yes

Circularity Check

0 steps flagged

No circularity in the proposed self-supervised framework

full rationale

The paper proposes Neighbor2Inverse as an empirical training procedure that adapts subsampling of noisy projections into pairs, performs separate reconstructions, and trains a denoising network in the image domain. This is validated through direct experiments on ROI PBI-CT and clinical CT data using external benchmarks and comparative metrics (CNR, spatial resolution, composite quality). No equations, fitted parameters, or derivation steps are presented that reduce predictions or results to the inputs by construction. The central claims rest on experimental outcomes rather than self-referential definitions or load-bearing self-citations.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The method depends on the domain assumption that subsampling creates independent noise realizations while preserving structure, drawn from the Neighbor2Neighbor principle.

axioms (1)

domain assumption Subsampling a noisy projection produces two variants with independent noise but identical underlying structure
Core premise enabling self-supervised pairs without high-dose references.

pith-pipeline@v0.9.0 · 5622 in / 1134 out tokens · 36467 ms · 2026-05-09T19:13:36.563612+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

41 extracted references · 32 canonical work pages

[1]

Phase objects in synchrotron radiation hard x-ray imaging

Peter Cloetens et al. “Phase objects in synchrotron radiation hard x-ray imaging”. In:Journal of Physics D: Applied Physics29.1 (1996), p. 133

1996
[2]

On the possibilities of x-ray phase contrast microimaging by coherent high-energy synchrotron radiation

Anatoly Snigirev et al. “On the possibilities of x-ray phase contrast microimaging by coherent high-energy synchrotron radiation”. In:Review of Scientific Instruments66.12 (1995), pp. 5486–5492

1995
[3]

Simultaneous phase and amplitude extraction from a single defocused image of a homogeneous object

D. Paganin et al. “Simultaneous phase and amplitude extraction from a single defocused image of a homogeneous object”. In:Journal of Microscopy206.1 (2002).ISSN: 00222720.DOI: 10.1046/j.1365-2818.2002.01010. x

work page doi:10.1046/j.1365-2818.2002.01010 2002
[4]

Refraction-enhanced x-ray imaging of mouse lung using synchrotron radiation source

N Yagi et al. “Refraction-enhanced x-ray imaging of mouse lung using synchrotron radiation source”. In:Medical Physics26.10 (1999), pp. 2190–2193

1999
[5]

Evaluating the feasibility of region-of-interest X-ray phase contrast imaging for lung cancer diagnostics

Lucy Costello et al. “Evaluating the feasibility of region-of-interest X-ray phase contrast imaging for lung cancer diagnostics”. In:Scientific Reports15 (June 2025).DOI:10.1038/s41598-025-04509-z

work page doi:10.1038/s41598-025-04509-z 2025
[6]

High-energy X-ray phase-contrast CT of an adult human chest phantom

Jannis N. Ahlers et al. “High-energy X-ray phase-contrast CT of an adult human chest phantom”. In:Scientific Reports15.1 (Aug. 2025).ISSN: 2045-2322.DOI: 10.1038/s41598-025-14956-3 .URL: http://dx.doi. org/10.1038/s41598-025-14956-3

work page doi:10.1038/s41598-025-14956-3 2025
[7]

In situ propagation-based lung computed tomography for large animal models

Lorenzo D’Amico et al. “In situ propagation-based lung computed tomography for large animal models”. In:Journal of Synchrotron Radiation32.6 (Oct. 2025), pp. 1511–1522.ISSN: 1600-5775.DOI: 10 . 1107 / s160057752500832x.URL:http://dx.doi.org/10.1107/S160057752500832X

work page doi:10.1107/s160057752500832x 2025
[8]

De Margerie-Mellon, C

C. De Margerie-Mellon, C. De Bazelaire, and E. De Kerviler.Image-guided biopsy in primary lung cancer: Why, when and how. 2016.DOI:10.1016/j.diii.2016.06.016

work page doi:10.1016/j.diii.2016.06.016 2016
[9]

Tumor seeding following CT- guided transthoracic needle biopsy in lung cancer. A case report

Thomas Melzer et al. “Tumor seeding following CT- guided transthoracic needle biopsy in lung cancer. A case report”. In:BMC Pulmonary Medicine23.1 (2023).ISSN: 14712466.DOI:10.1186/s12890-023-02712-0

work page doi:10.1186/s12890-023-02712-0 2023
[10]

Biopsy frequency and complications among lung cancer patients in the United States

Yichen Zhang et al. “Biopsy frequency and complications among lung cancer patients in the United States”. In: Lung Cancer Management9.4 (2020).ISSN: 17581974.DOI:10.2217/lmt-2020-0022

work page doi:10.2217/lmt-2020-0022 2020
[11]

Reid, and Silvio Savarese

Shi Guo et al. “Toward convolutional blind denoising of real photographs”. In:Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition. V ol. 2019-June. 2019.DOI:10.1109/CVPR. 2019.00181

work page doi:10.1109/cvpr 2019
[12]

Noise2Noise: Learning image restoration without clean data

Jaakko Lehtinen et al. “Noise2Noise: Learning image restoration without clean data”. In:35th International Conference on Machine Learning, ICML 2018. V ol. 7. 2018. 11

2018
[13]

2021 , url =

Tao Huang et al. “Neighbor2Neighbor: Self-supervised denoising from single noisy images”. In:Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition. 2021.DOI: 10.1109/ CVPR46437.2021.01454

work page arXiv 2021
[14]

Vision transformers are parameter- efficient audio-visual learners

Youssef Mansour and Reinhard Heckel. “Zero-Shot Noise2Noise: Efficient Image Denoising without any Data”. In:Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition. V ol. 2023-June. 2023.DOI:10.1109/CVPR52729.2023.01347

work page doi:10.1109/cvpr52729.2023.01347 2023
[15]

A., Pelt, D

Allard Adriaan Hendriksen, Daniel Maria Pelt, and K. Joost Batenburg. “Noise2Inverse: Self-Supervised Deep Convolutional Denoising for Tomography”. In:IEEE Transactions on Computational Imaging6 (2020).ISSN: 23339403.DOI:10.1109/TCI.2020.3019647

work page doi:10.1109/tci.2020.3019647 2020
[16]

Proj2Proj: self-supervised low-dose CT reconstruction

Mehmet Ozan Unal, Metin Ertas, and Isa Yildirim. “Proj2Proj: self-supervised low-dose CT reconstruction”. In: PeerJ Computer Science10 (2024).ISSN: 23765992.DOI:10.7717/peerj-cs.1849

work page doi:10.7717/peerj-cs.1849 2024
[17]

Shadows can be

Zejin Wang et al. “Blind2Unblind: Self-Supervised Image Denoising with Visible Blind Spots”. In:Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition. V ol. 2022-June. 2022. DOI:10.1109/CVPR52688.2022.00207

work page doi:10.1109/cvpr52688.2022.00207 2022
[18]

Yipeng Sun et al.Filter2Noise: A Framework for Interpretable and Zero-Shot Low-Dose CT Image Denoising
[19]

arXiv:2504.13519 [eess.IV].URL:https://arxiv.org/abs/2504.13519

work page arXiv
[20]

SDCNN: Self-Supervised Disentangled Convolutional Neural Network for Low-Dose CT Denoising

Yuhang Liu et al. “SDCNN: Self-Supervised Disentangled Convolutional Neural Network for Low-Dose CT Denoising”. In:IEEE Transactions on Instrumentation and Measurement74 (2025).ISSN: 15579662.DOI: 10.1109/TIM.2024.3502758

work page doi:10.1109/tim.2024.3502758 2025
[21]

Mettle: Meta-token learning for memory-efficient audio-visual adaptation.IEEE Trans

Qing Ma et al. “Pixel2Pixel: A Pixelwise Approach for Zero-Shot Single Image Denoising”. In:IEEE Transac- tions on Pattern Analysis and Machine Intelligence47.6 (2025).ISSN: 19393539.DOI: 10.1109/TPAMI.2025. 3546870

work page doi:10.1109/tpami.2025 2025
[22]

Detectors for the imaging and medical beam line at the australian synchrotron

C. Hall et al. “Detectors for the imaging and medical beam line at the australian synchrotron”. In:Journal of Instrumentation. V ol. 8. 6. June 2013.DOI:10.1088/1748-0221/8/06/C06011

work page doi:10.1088/1748-0221/8/06/c06011 2013
[23]

The RSNA Pulmonary Embolism CT Dataset

Errol Colak et al. “The RSNA Pulmonary Embolism CT Dataset”. In:Radiology: Artificial Intelligence3.2 (2021).ISSN: 26386100.DOI:10.1148/ryai.2021200254

work page doi:10.1148/ryai.2021200254 2021
[24]

Torchradon: Fast differentiable routines for computed tomography,

Matteo Ronchetti. “TorchRadon: Fast Differentiable Routines for Computed Tomography”. In:arXiv preprint arXiv:2009.14788(2020). eprint:arXiv:2009.14788

work page arXiv 2009
[25]

SYRMEP Tomo Project: a graphical user interface for customizing CT reconstruction workflows

Francesco Brun et al. “SYRMEP Tomo Project: a graphical user interface for customizing CT reconstruction workflows”. In:Advanced Structural and Chemical Imaging3.1 (2017).ISSN: 21980926.DOI: 10.1186/s40679- 016-0036-8

work page doi:10.1186/s40679- 2017
[26]

Superior techniques for eliminating ring artifacts in X-ray micro-tomography

Nghia T. V o, Robert C. Atwood, and Michael Drakopoulos. “Superior techniques for eliminating ring artifacts in X-ray micro-tomography”. In:Optics Express26.22 (2018).ISSN: 10944087.DOI:10.1364/oe.26.028396

work page doi:10.1364/oe.26.028396 2018
[27]

Fast iterative reconstruction of data in full interior tomography

F. Arcadu, F. Marone, and M. Stampanoni. “Fast iterative reconstruction of data in full interior tomography”. In: Journal of Synchrotron Radiation. V ol. 24. 1. 2017.DOI:10.1107/S1600577516015794

work page doi:10.1107/s1600577516015794 2017
[28]

Principles of Statistical Radiophysics

S. M. Rytov et al. “ Principles of Statistical Radiophysics ”. In:Physics Today42.12 (1989).ISSN: 0031-9228. DOI:10.1063/1.2811249

work page doi:10.1063/1.2811249 1989
[29]

DETERMINISTIC PHASE RETRIEV AL: A GREEN’S FUNCTION SOLUTION

Michael Reed Teague. “DETERMINISTIC PHASE RETRIEV AL: A GREEN’S FUNCTION SOLUTION.” In: Journal of the Optical Society of America73.11 (1983).ISSN: 00303941.DOI:10.1364/JOSA.73.001434

work page doi:10.1364/josa.73.001434 1983
[30]

Improved training of wasserstein GANs

Ishaan Gulrajani et al. “Improved training of wasserstein GANs”. In:Advances in Neural Information Processing Systems. V ol. 2017-December. 2017

2017
[31]

Dabov, A

Kostadin Dabov et al. “Image denoising by sparse 3-D transform-domain collaborative filtering”. In:IEEE Transactions on Image Processing16.8 (2007).ISSN: 10577149.DOI:10.1109/TIP.2007.901238

work page doi:10.1109/tip.2007.901238 2007
[32]

Collaborative Filtering of Correlated Noise: Exact Transform- Domain Variance for Improved Shrinkage and Patch Matching

Ymir Makinen, Lucio Azzari, and Alessandro Foi. “Collaborative Filtering of Correlated Noise: Exact Transform- Domain Variance for Improved Shrinkage and Patch Matching”. In:IEEE Transactions on Image Processing29 (2020).ISSN: 19410042.DOI:10.1109/TIP.2020.3014721

work page doi:10.1109/tip.2020.3014721 2020
[33]

An Algorithm for Total Variation Minimization and Applications

Antonin Chambolle. “An Algorithm for Total Variation Minimization and Applications”. In:Journal of Mathe- matical Imaging and Vision. V ol. 20. 1-2. 2004.DOI:10.1023/B:JMIV.0000011321.19549.88

work page doi:10.1023/b:jmiv.0000011321.19549.88 2004
[34]

E., et al

Pauli Virtanen et al. “SciPy 1.0: fundamental algorithms for scientific computing in Python”. In:Nature Methods 17.3 (2020).ISSN: 15487105.DOI:10.1038/s41592-019-0686-2

work page doi:10.1038/s41592-019-0686-2 2020
[35]

U-Net: convolutional networks for biomedical image segmentation , booktitle =

Olaf Ronneberger, Philipp Fischer, and Thomas Brox. “U-net: Convolutional networks for biomedical image segmentation”. In:Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). V ol. 9351. 2015.DOI:10.1007/978-3-319-24574-4{\_}28

work page doi:10.1007/978-3-319-24574-4 2015
[36]

Efficientnetv2: Smaller models and faster training,

Mingxing Tan and Quoc V . Le. “EfficientNetV2: Smaller Models and Faster Training”. In:CoRRabs/2104.00298 (2021). arXiv:2104.00298.URL:https://arxiv.org/abs/2104.00298. 12

work page arXiv 2021
[37]

PyTorch: An imperative style, high-performance deep learning library

Adam Paszke et al. “PyTorch: An imperative style, high-performance deep learning library”. In:Advances in Neural Information Processing Systems. V ol. 32. 2019

2019
[38]

Version 1.4

William Falcon and The PyTorch Lightning team.PyTorch Lightning. Version 1.4. Mar. 2019.DOI: 10.5281/ zenodo.3828935.URL:https://github.com/Lightning-AI/lightning

2019
[39]

Adam: A method for stochastic optimization

Diederik P. Kingma and Jimmy Lei Ba. “Adam: A method for stochastic optimization”. In:3rd International Conference on Learning Representations, ICLR 2015 - Conference Track Proceedings. 2015

2015
[40]

, month = mar, year =

Leslie N. Smith. “Cyclical learning rates for training neural networks”. In:Proceedings - 2017 IEEE Winter Conference on Applications of Computer Vision, WACV 2017. 2017.DOI:10.1109/WACV.2017.58

work page doi:10.1109/wacv.2017.58 2017
[41]

2025.DOI: 10.5281/zenodo.15469773.URL: https://github.com/jnahlers/ fileswell

Jannis Ahlers.Fileswell. 2025.DOI: 10.5281/zenodo.15469773.URL: https://github.com/jnahlers/ fileswell. 13

work page doi:10.5281/zenodo.15469773.url: 2025