Pith. sign in

REVIEW 3 major objections 5 minor 28 references

HyperSORT: Self-Organising Robust Training with hyper-networks

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Jointly learned per-sample latent vectors and a hyper-network map a medical segmentation training set into clusters that correspond to annotation styles, with outlier norms flagging erroneous labels.

desk verdict A novel self-organizing hypernetwork for segmentation, but the error-detection claim is not yet tested; needs a discrimination experiment. read the letter →

arxiv 2506.21430 v1 pith:FPMMO6PU submitted 2025-06-26 cs.CV

Samuel Joutard , Marijn Stollenga , Marc Balle Sanchez , Mohammad Farid Azampour , Raphael Prevost This is my paper · ORCID
classification cs.CV
keywords hyper-networksrobusttrainingself-organisingmedicalimagesegmentationnoisylabelsdatasetcurationlatentspacemappingannotationbias
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

HyperSORT's hypothesis is that the labeling process behind a medical imaging dataset can be modeled by a hidden per-sample variable: each image/label pair is generated by an unknown oracle that depends on a latent annotation-style vector. The paper shows that by learning, for every training sample, a low-dimensional latent vector and a hyper-network that maps that vector to the weights of a UNet, the training set organizes itself into clusters that correspond to distinct annotation behaviors, while an L1 penalty keeps the dominant style near the origin and pushes unusual or erroneous cases outward. On a synthetically perturbed AMOS liver dataset the learned latent space separates normal, eroded, and dilated labels along one axis; on TotalSegmentator V1 the latent norm and isolation measures correlate with the amount of correction applied in V2, outperforming two label-quality baselines. The UNets generated from cluster centroids reach Dice scores comparable to reference models, and the different clusters produce deliberately different segmentations of the same case, which a human can choose among. The aim is to show that one training procedure can simultaneously build a competitive segmenter, expose systematic annotation biases, and flag erroneous labels.

What carries the argument

The load-bearing object is the pair (latent vector, hyper-network): a trainable vector $\bar\lambda_n\in\mathbb{R}^2$ per training sample and a small fully connected hyper-network that emits the entire parameter tensor of a 3D UNet from that vector. The L1 regularization term $L_{\rm reg}(\bar\lambda_n)$ is what turns an unconstrained conditioning variable into a self-organizing map, because it forces the main annotation mode to sit at the origin and makes Euclidean distance from the origin a proxy for annotation deviation. The 2D latent choice is not incidental: it is what allows the clusters to be visualized and inspected, and the paper argues that the learned ordering (eroded/normal/dilated along one axis) is meaningful, with the orthogonal axis absorbing secondary variability such as challenging anatomy. The hyper-network also enables synergistic learning, since all styles share the same generating function and can transfer strength between clusters.

What would settle it

Train HyperSORT on a dataset with known rater identities (or with clean labels) and inspect the latent map: if same-rater samples do not cluster, or if a clean dataset still produces large high-norm outlier clusters, the claim that the latent geometry is shaped by annotation style would be contradicted. Alternatively, an ablation that randomizes which samples are assigned to which latent vector should destroy the cluster-error correlation if the mechanism is real.

Watch

Extended reading notes

Core claim

The central claim is that the empirical distribution of a training set's annotation variability can be captured by jointly optimizing per-sample latent vectors $\bar\lambda_n$ and a hyper-network $H_\beta$ that predicts the primary UNet's parameters as $U_{H_\beta(\bar\lambda_n)}$, with an L1 regularizer pushing the latent vectors toward zero (Eq. 1). Because the origin is the only privileged point, it becomes the learned 'normative' annotation style; samples whose labels deviate from that norm end up farther from zero, and the directions of the latent space become meaningful axes of annotation variation. The evidence is that the AMOS latent map separates the injected eroded and dilated styles in exactly the order of boundary tightness, and that on TotalSegmentator the latent-norm and mean-distance predictors obtain Spearman correlations of 0.2166 and 0.1723 with the number of voxels changed between V1 and V2, beating the Quality Sentinel and Test-Dice baselines. The paper further claims that hyper-network conditioning lets the different styles train synergistically rather than as independent models, so that even small clusters yield generalizing UNets with slightly higher Dice diversity across predictions.

Load-bearing premise

The error-detection story relies on the L1 penalty making the zero vector the single normative annotation style, so that distance from the origin measures how much a label deviates; if the dominant annotation mode is off-center or the main latent axes encode anatomy or acquisition differences rather than label errors, the cluster interpretation and the reported correlations no longer follow.

Editorial extensions

If this is right

  • A single HyperSORT run produces several usable segmenters (one per latent cluster centroid), so a clinician can choose the annotation style that best matches a target dataset instead of committing to one model.
  • Error discovery becomes a byproduct of training: latent-norm and mean-distance-to-all-samples rank which training cases to review, and the ranking beat the compared label-quality baselines on TotalSegmentator.
  • Cluster-specific UNets can be applied back to the training set to correct erroneous labels and generate pseudo-labels, supporting bootstrapping and dataset re-annotation workflows.
  • Because the learned UNet solutions are more diverse in their predictions (about twice the per-case Dice standard deviation of random seeds), the framework explores the segmentation solution space more thoroughly and improves best-of-$k$ outcomes.

Reading between the lines

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

  • A natural next test is to scale the latent dimension beyond 2: the authors note that higher-dimensional latents may give a more homogeneous relationship between norm and annotation deviation, which would strengthen the error-detection metric.
  • Because the map also distances under-represented but correctly labeled cohorts, HyperSORT could be used prospectively to sample a training set for diversity before labeling, not only after errors have occurred.
  • The TotalSegmentator correlation is measured against V1-to-V2 corrections, which are themselves an imperfect pseudo ground truth; a stronger validation would compare latent outliers against independent multi-rater re-annotations of the same scans.
  • One speculative but testable extension: if latent axes are aligned to known rater identity, the method could quantify inter-rater agreement and reveal which style differences matter for a given segmentation task.
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. The manuscript introduces HyperSORT, a framework in which a hyper-network predicts the parameters of a 3D UNet from a per-sample latent vector. The latent vectors and hyper-network parameters are trained jointly on a segmentation loss plus an L1 regularizer on the latent vectors, with the intended effect that the main annotation style is centered near the origin and atypical or erroneous samples appear as outliers or clusters. The method is evaluated on a synthetic perturbed AMOS liver dataset and on TotalSegmentator V1 versus V2, using V2 corrections as a pseudo ground truth. The authors claim that HyperSORT produces competitive segmentation UNets, a structured mapping of the training set that reveals systematic biases, and latent-space features that rank the severity of label errors. The paper includes a public code repository and visual analyses of the learned latent spaces.

Significance. If the central claims hold, HyperSORT would be a useful tool for medical image segmentation dataset curation, combining robust training with an interpretable latent map. The paper has several strengths: the controlled synthetic perturbation experiment on AMOS provides a clear proof of concept; the TotalSegmentator V1-to-V2 comparison uses an external pseudo ground truth; the method does not derive its constants from the fitted parameters under test; and the code is publicly released. However, the central claim that the method 'allows the identification of erroneous samples' is currently under-supported: the reported quantitative evidence is a rank correlation with correction magnitude among already-corrected cases, not a discrimination test between erroneous and correct cases. The manuscript's own Discussion (Sec. 4.3) acknowledges that distinguishing 'bad labels' from 'challenging correct labels' remains an open problem, which is exactly the boundary of the claimed capability. The significance of the contribution is therefore contingent on adding a proper discrimination evaluation and on validating the assumed normative role of the latent-space origin.

major comments (3)
  1. [Sec. 4.2] The quantitative evidence for the statement that the latent map can be used to identify erroneous cases does not actually test identification. The pseudo-ground-truth analysis considers only cases with at least one voxel changed between TS-V1 and TS-V2, and the reported Spearman correlations (0.2166 for ||lambda_n|| and 0.1723 for mean distance) rank the severity of known corrections among already-corrected cases. No comparison is made between corrected and non-corrected cases over the full 1204-case set, and no AUC, precision-recall, or confusion-matrix results are reported. Since the abstract and conclusion claim 'identification of erroneous samples', a discrimination test is load-bearing. Without it, the observed correlations are compatible with the alternative explanation that the latent norm measures atypicality in general, including rare anatomy, acquisition artifacts, or hard cases, which is precisely the confound acknowledged in Sec. 4.3. I request a ROC/AUC or precision-recall analysis over all training samples using V1-to-V2 correction as the label, with the same baselines, and a report of the latent-norm distributions for corrected versus non-corrected cases.
  2. [Sec. 3, Eq. (1)] The interpretation of ||lambda_n|| as label-error severity rests on the assumption that the L1 regularizer places the main annotation style at the origin, but Eq. (1) only pushes lambda_n toward zero; nothing in the formulation guarantees that the dominant empirical mode of the annotation-style distribution coincides with the zero vector in the 2D latent space, especially with the hyper-network output capped by tanh(x)*5. If the main style is off-center, or if the dominant latent axes encode anatomy or acquisition differences rather than label quality, then the norm-based and mean-distance predictors lose their claimed meaning. Please provide direct validation of the normative-origin assumption, for example by reporting the fraction of known-normal AMOS cases in the zero-centered cluster, the cluster centroids and their Dice agreement with TS-V2 corrections, and the latent-norm distributions for known-correct versus known-corrected TS cases. This is load-bearing for the error-detection claim and for the inference choice of lambda = 0 as the normative style.
  3. [Sec. 4.2, Table 1] The 'human-in-the-loop' best-out-of-five comparison is not a valid evaluation of a deployable inference strategy: selecting the best Dice per test case requires access to the test ground truth, so the reported 'systematically better predictions' (p <= 1e-5) mainly demonstrate that the five HyperSORT UNets are more diverse (per-case Dice std 0.5 versus 0.2), not that a user can attain that best-case performance in practice. Please either evaluate a feasible selection mechanism, such as always using the zero-cluster UNet or an unsupervised confidence score, or explicitly label the best-out-of-five numbers as an oracle upper bound rather than a human-in-the-loop result.
minor comments (5)
  1. [Sec. 4.2] The number of cases used in the Spearman correlation analysis is not reported; please state how many of the 1204 TotalSegmentator cases had at least one voxel changed and how many had more than 10000 voxels changed.
  2. [Sec. 3 and Sec. 4.2] The cluster structure is central to the reported experiments, but the text does not specify how the five clusters in Figure 3 are obtained, whether by a clustering algorithm, by visual inspection, or by manual selection; please describe the cluster-extraction procedure or state that the ellipses are illustrative, and report cluster sizes.
  3. [Sec. 3, Eq. (1)] The exact form of Lreg is not specified beyond 'L1-norm regularization'; please give the coefficient and the precise expression, since this term determines the degree to which the main mode is pulled toward the origin.
  4. [Sec. 4.2] The claim that the red cluster's UNet 'explains the poorer performance of that cluster's UNet on the CT-1K dataset' is made from a small number of visually inspected cases; consider quantifying the cluster-level error rate against V2 corrections to support the systematic-error interpretation.
  5. [Fig. 2] The phrase 'most eccentric cases' in the caption is not defined; please specify the criterion used to select the displayed examples.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: external benchmarks carry the validation and the one self-citation is context only.

full rationale

HyperSORT's core objects are the per-sample latent vectors and the hypernetwork, jointly fitted by Eq. (1) as a minimization over the hypernetwork parameters and latent collection of the segmentation loss plus an L1 regularization term. The downstream claims are not derived from those fitted parameters by construction: the AMOS experiment evaluates whether latent clusters separate known injected erosion/dilation perturbations; the TotalSegmentator experiment correlates latent-norm and isolation predictors against independently published V1-to-V2 voxel changes; and CT-1K provides an external segmentation benchmark. None of these targets appears in the training objective, so the reported correlations and cluster correspondences are empirical findings rather than forced fits. The sole overlapping-author citation, ref. [15] (HyperSpace), appears in Related Work as an example of spacing-conditioned hypernetworks and is not used to justify HyperSORT's novelty or its error-detection claim. Section 4.3 candidly notes that distinguishing 'bad labels' from 'challenging correct labels' remains open; this is a limitation in the strength of the error-detection evidence, not a circular step. No equation reduces to its own input and no fitted parameter is renamed as an independent prediction.

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

The central claims depend on a small set of hand-chosen architecture and regularization choices, plus assumptions that the latent space encodes annotation style and that V1-to-V2 corrections are true errors.

free parameters (5)
  • 2-dimensional latent vector dimension = 2
    Chosen for visualization; no ablation on higher dimensions; affects whether clusters map linearly to annotation style.
  • L1 regularization weight on latent vectors = not specified in text
    Controls trade-off between per-sample fitting and compactness; not reported, so reproducibility and sensitivity unknown.
  • Number of latent clusters (5) = 5
    Selected by visual inspection of Figure 3; not obtained via clustering algorithm, so cluster assignment is subjective.
  • Hypernetwork architecture = FC 3x50, ReLU
    Chosen without ablations; capacity affects ability to fit diverse UNet parameters.
  • Parameter capping constant = 5 (tanh(x)*5)
    Introduced to stabilize training; effect on final latent structure not analyzed.
assumptions (4)
  • domain assumption The annotation process can be modeled as a deterministic oracle Ω(I, λ) with additive iid noise.
    Section 3 defines the labeling process; this is a modeling assumption that may not hold when annotation errors are systematic and non-additive.
  • ad hoc to paper The L1 regularization toward the origin makes the main annotation mode located near the zero vector, so λ=0 is the normative style.
    Section 3 states this; inference at λ=0 and latent-norm error detection both rely on it.
  • domain assumption Euclidean distance in 2D latent space reflects annotation style difference.
    Used for the 'isolation' predictor in Section 4.2; no validation that the latent metric aligns with label-error severity.
  • domain assumption Corrections from TotalSegmentator V1 to V2 are an accurate proxy for labeling errors.
    Section 4.2 uses voxel changes between V1 and V2 as pseudo ground truth for error detection; this assumes all changes are corrections of errors.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HyperSORT: Self-Organising Robust Training with hyper-networks." pith.science (2026). https://pith.science/paper/FPMMO6PU

@misc{pith2026250621430,
  author       = {Pith},
  title        = {Pith review of: HyperSORT: Self-Organising Robust Training with hyper-networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FPMMO6PU}},
  note         = {Machine review of arXiv:2506.21430}
}
read the original abstract

Medical imaging datasets often contain heterogeneous biases ranging from erroneous labels to inconsistent labeling styles. Such biases can negatively impact deep segmentation networks performance. Yet, the identification and characterization of such biases is a particularly tedious and challenging task. In this paper, we introduce HyperSORT, a framework using a hyper-network predicting UNets' parameters from latent vectors representing both the image and annotation variability. The hyper-network parameters and the latent vector collection corresponding to each data sample from the training set are jointly learned. Hence, instead of optimizing a single neural network to fit a dataset, HyperSORT learns a complex distribution of UNet parameters where low density areas can capture noise-specific patterns while larger modes robustly segment organs in differentiated but meaningful manners. We validate our method on two 3D abdominal CT public datasets: first a synthetically perturbed version of the AMOS dataset, and TotalSegmentator, a large scale dataset containing real unknown biases and errors. Our experiments show that HyperSORT creates a structured mapping of the dataset allowing the identification of relevant systematic biases and erroneous samples. Latent space clusters yield UNet parameters performing the segmentation task in accordance with the underlying learned systematic bias. The code and our analysis of the TotalSegmentator dataset are made available: https://github.com/ImFusionGmbH/HyperSORT

Figures

Figures reproduced from arXiv: 2506.21430 by the authors.

Figure 1
Figure 1. Overview of HyperSORT. The hyper-network Hβ generates the UNet parame￾ters θ from a conditioning vector λ¯n specific to a training sample In, Sn. HyperSORT tackles this problem by modeling the annotation process with an additional hidden variable. As such, this hidden variable can parameterize differ￾ences between raters or annotation errors. A hyper-network [10] conditions the segmentation UNet [20] behavior on thi… view at source ↗
Figure 2
Figure 2. (Left) Obtained AMOS latent space. The most eccentric cases of the −→0 cluster are challenging cases (from left to right): incomplete liver segmentation, liver tissue heterogeneity, and abnormal abdominal anatomy where the gallbladder and stomach are shown for reference. (Right) Inference using the −→0 latent on images from the erosion and dilation clusters. 4 Experiments and Results HyperSORT is agnostic to the cho… view at source ↗
Figure 3
Figure 3. (Left) Obtained TS latent space. Blue and green dots respectively correspond to non-corrected and corrected cases from V1 to V2. Green dots’ radii are proportional to the number of voxels modified. 5 visual clusters are highlighted with colored ellipses. (Right) Slices from 3 cases belonging to the cluster that were not modified between V1 and V2. Purple arrows highlight erroneous liver labeling. Corrective predicti… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 22 canonical work pages

  1. [1]

    Adaptive Computation and Ma- chine Learning series, MIT Press (2024)

    Bach, F.: Learning Theory from First Principles. Adaptive Computation and Ma- chine Learning series, MIT Press (2024)

    work page 2024

  2. [2]

    In: Shen, D., Liu, T., Peters, T.M., Staib, L.H., Essert, C., Zhou, S., Yap, P.T., Khan, A

    Baumgartner, C.F., Tezcan, K.C., Chaitanya, K., Hötker, A.M., Muehlematter, U.J., Schawkat, K., Becker, A.S., Donati, O., Konukoglu, E.: Phiseg: Capturing uncertainty in medical image segmentation. In: Shen, D., Liu, T., Peters, T.M., Staib, L.H., Essert, C., Zhou, S., Yap, P.T., Khan, A. (eds.) Medical Image Com- puting and Computer Assisted Intervention...

    work page 2019

  3. [3]

    In: Medical Imaging with Deep Learn- ing (2024), https://openreview.net/forum?id=sfjgmuvLS7

    Billot, B., Dey, N., Turk, E.A., Grant, E., Golland, P.: Network conditioning for synergistic learning on partial annotations. In: Medical Imaging with Deep Learn- ing (2024), https://openreview.net/forum?id=sfjgmuvLS7

    work page 2024

  4. [4]

    Medical Image Analysis71, 102062 (2021)

    Budd,S.,Robinson,E.C.,Kainz,B.:Asurveyonactivelearningandhuman-in-the- loop deep learning for medical image analysis. Medical Image Analysis71, 102062 (2021). https://doi.org/https://doi.org/10.1016/j.media.2021.102062, https://www.sciencedirect.com/science/article/pii/S1361841521001080 10 S. Joutard et al

    work page arXiv 2021

  5. [5]

    Chen, J., Ramanathan, V., Xu, T., Martel, A.L.: Detecting noisy labels with repeated cross-validations LNCS 15010 (October 2024)

    work page 2024

  6. [6]

    Quality Sentinel: Estimating Label Quality and Errors in Medical Segmentation Datasets

    Chen, Y., Zhou, Z., Yuille, A.L.: Quality sentinel: Estimating label quality and errors in medical segmentation datasets. CoRRabs/2406.00327 (2024), https: //doi.org/10.48550/arXiv.2406.00327

    work page Pith review arXiv doi:10.48550/arxiv.2406.00327 2024

  7. [7]

    Frontiers of Computer Science19(9), 199705 (Jan 2025)

    Dong, W., Du, B., Xu, Y.: Shape-intensity knowledge distillation for robust medical image segmentation. Frontiers of Computer Science19(9), 199705 (Jan 2025). https://doi.org/10.1007/s11704-024-40462-2, https://doi.org/10. 1007/s11704-024-40462-2

    work page doi:10.1007/s11704-024-40462-2 2025

  8. [8]

    European Radiology Experimental 8(1), 11 (Feb 2024)

    Galbusera, F., Cina, A.: Image annotation and curation in radiology: an overview for machine learning practitioners. European Radiology Experimental 8(1), 11 (Feb 2024). https://doi.org/10.1186/s41747-023-00408-y, https://doi.org/ 10.1186/s41747-023-00408-y

    work page doi:10.1186/s41747-023-00408-y 2024

Show all 28 references
  1. [9]

    In: Greenspan, H., Madab- hushi, A., Mousavi, P., Salcudean, S., Duncan, J., Syeda-Mahmood, T., Taylor, R

    Gonzalez-Jimenez, A., Lionetti, S., Gottfrois, P., Gröger, F., Pouly, M., Navarini, A.A.: Robust t-loss for medical image segmentation. In: Greenspan, H., Madab- hushi, A., Mousavi, P., Salcudean, S., Duncan, J., Syeda-Mahmood, T., Taylor, R. (eds.) Medical Image Computing and...

  2. [10]

    In: International Conference on Learning Representations (2017),https://openreview.net/forum?id=rkpACe1lx

    Ha, D., Dai, A.M., Le, Q.V.: Hypernetworks. In: International Conference on Learning Representations (2017),https://openreview.net/forum?id=rkpACe1lx

    2017

  3. [11]

    Machine Learn- ing for Biomedical Imaging1, 1–30 (2022).https://doi.org/10.59275/j.melba

    Hoopes, A., Hoffmann, M., Greve, D.N., Fischl, B., Guttag, J., Dalca, A.: Learn- ing the effect of registration hyperparameters with hypermorph. Machine Learn- ing for Biomedical Imaging1, 1–30 (2022).https://doi.org/10.59275/j.melba. 2022-74f1

    2022 doi

  4. [12]

    Pattern Recognition 158, 111028 (2025).https://doi.org/https://doi.org/10.1016/j

    Iqbal, S., Khan, T.M., Naqvi, S.S., Naveed, A., Meijering, E.: Tbconvl-net: A hybrid deep learning architecture for robust medical image segmentation. Pattern Recognition 158, 111028 (2025).https://doi.org/https://doi.org/10.1016/j. patcog.2024.111028, https://www.sciencedirec...

    2025

  5. [13]

    Nature Methods 18(2), 203–211 (Feb 2021)

    Isensee, F., Jaeger, P.F., Kohl, S.A.A., Petersen, J., Maier-Hein, K.H.: nnu-net: a self-configuring method for deep learning-based biomedical image segmenta- tion. Nature Methods 18(2), 203–211 (Feb 2021). https://doi.org/10.1038/ s41592-020-01008-z, https://doi.org/10.1038/s...

    2021 doi

  6. [14]

    arXiv preprint arXiv:2206.08023 (2022)

    Ji, Y., Bai, H., Yang, J., Ge, C., Zhu, Y., Zhang, R., Li, Z., Zhang, L., Ma, W., Wan, X., et al.: Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation. arXiv preprint arXiv:2206.08023 (2022)

    2022 arXiv

  7. [15]

    In: proceedings of Medical Image Computing and Computer Assisted Intervention – MICCAI 2024

    Joutard, S., Pietsch, M., Prevost, R.: HyperSpace: Hypernetworks for spacing- adaptive image segmentation . In: proceedings of Medical Image Computing and Computer Assisted Intervention – MICCAI 2024. vol. LNCS 15009. Springer Na- ture Switzerland (October 2024)

    2024

  8. [16]

    arXiv preprint arXiv:2307.05080 (2023)

    Lad, V., Mueller, J.: Estimating label quality and errors in semantic segmentation data via any model. arXiv preprint arXiv:2307.05080 (2023)

    2023 arXiv

  9. [17]

    Ma, J., Zhang, Y., Gu, S., Zhang, Y., Zhu, C., Wang, Q., Liu, X., An, X., Ge, C., Cao, S., Zhang, Q., Liu, S., Wang, Y., Li, Y., Wang, C., He, J., Yang, X.: Abdomenct-1k: Is abdominal organ segmentation a solved problem? IEEE Transactions on Pattern Analysis and Machine Intell...

    2020

  10. [18]

    CoRRabs/1906.03815 (2019), http://arxiv.org/abs/1906.03815 HyperSORT: Self-Organising Robust Training with hyper-networks 11

    Mirikharaji,Z.,Yan,Y.,Hamarneh,G.:Learningtosegmentskinlesionsfromnoisy annotations. CoRRabs/1906.03815 (2019), http://arxiv.org/abs/1906.03815 HyperSORT: Self-Organising Robust Training with hyper-networks 11

    2019 arXiv

  11. [19]

    In: de Bruijne, M., Cattin, P.C., Cotin, S., Padoy, N., Speidel, S., Zheng, Y., Essert, C

    Mok, T.C.W., Chung, A.C.S.: Conditional deformable image registration with con- volutional neural network. In: de Bruijne, M., Cattin, P.C., Cotin, S., Padoy, N., Speidel, S., Zheng, Y., Essert, C. (eds.) MICCAI 2021. pp. 35–45. Springer Inter- national Publishing, Cham (2021)

    2021

  12. [20]

    In: Navab, N., Hornegger, J., Wells, W.M., Frangi, A.F

    Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed- ical image segmentation. In: Navab, N., Hornegger, J., Wells, W.M., Frangi, A.F. (eds.) MICCAI 2015 (2015)

    2015

  13. [21]

    npj Digital Medicine6(1), 26 (Feb 2023)

    Sylolypavan, A., Sleeman, D., Wu, H., Sim, M.: The impact of inconsistent human annotations on ai driven clinical decision making. npj Digital Medicine6(1), 26 (Feb 2023). https://doi.org/10.1038/s41746-023-00773-3, https://doi.org/ 10.1038/s41746-023-00773-3

    2023 doi

  14. [22]

    In: Medical Imaging with Deep Learning (2022),https://openreview.net/forum? id=C4B46ZS7MSB

    S,erban Vădineanu, Pelt, D., Dzyubachyk, O., Batenburg, J.: An analysis of the impact of annotation errors on the accuracy of deep learning for cell segmentation. In: Medical Imaging with Deep Learning (2022),https://openreview.net/forum? id=C4B46ZS7MSB

    2022

  15. [23]

    Radiology: Arti- ficial Intelligence 5(5), e230024 (2023)

    Wasserthal, J., Breit, H.C., Meyer, M.T., Pradella, M., Hinck, D., Sauter, A.W., Heye, T., Boll, D.T., Cyriac, J., Yang, S., Bach, M., Segeroth, M.: Totalsegmenta- tor: Robust segmentation of 104 anatomic structures in ct images. Radiology: Arti- ficial Intelligence 5(5), e230...

    2023 doi

  16. [24]

    Journal of the American Medical Informatics Association 31(7), 1596–1607 (05 2024)

    Wei, Y., Deng, Y., Sun, C., Lin, M., Jiang, H., Peng, Y.: Deep learning with noisy labels in medical prediction problems: a scoping review. Journal of the American Medical Informatics Association 31(7), 1596–1607 (05 2024). https://doi.org/ 10.1093/jamia/ocae108, https://doi.o...

    2024 doi

  17. [25]

    Radiology 295(1), 4–15 (2020)

    Willemink, M.J., Koszek, W.A., Hardell, C., Wu, J., Fleischmann, D., Har- vey, H., Folio, L.R., Summers, R.M., Rubin, D.L., Lungren, M.P.: Prepar- ing medical imaging data for machine learning. Radiology 295(1), 4–15 (2020). https://doi.org/10.1148/radiol.2020192224, https://d...

    2020 doi

  18. [26]

    NeurIPS (2020)

    Zhang,L.,Tanno,R.,Xu,M.C.,Jacob,J.,Ciccarelli,O.,Barkhof,F.,C.Alexander, D.: Disentangling human error from the ground truth in segmentation of medical images. NeurIPS (2020)

    2020

  19. [27]

    In: Proceedings of the 32nd International Conference on Neural Information Processing Systems

    Zhang, Z., Sabuncu, M.R.: Generalized cross entropy loss for training deep neural networks with noisy labels. In: Proceedings of the 32nd International Conference on Neural Information Processing Systems. p. 8792–8802. NIPS’18, Curran Associates Inc., Red Hook, NY, USA (2018)

    2018

  20. [2023]

    pp. 714–724. Springer Nature Switzerland, Cham (2023)

    2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.