Pith. sign in

REVIEW 3 major objections 4 minor 62 references

Out-of-distribution data supervision towards biomedical semantic segmentation

T0 review · 3 major / 4 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 The paper claims Med-OoD, a purely data-centric method, mines negative OoD samples from the training set itself to cut foreground/background misclassification in medical segmentation, and that OoD-only training reaches 76.1% mIoU on Lizard.

desk verdict The 76.1% OOD-only result contradicts the paper's own loss; the combined-training idea is modest but under-validated. read the letter →

arxiv 2507.12105 v1 pith:KBXKHPE2 submitted 2025-07-16 cs.CV

Yiquan Gao , Duohui Xu This is my paper
classification cs.CV
keywords out-of-distributiondatabiomedicalsemanticsegmentationdata-centriclearningpositive-negativesampleratioLizarddatasetnegativesupervisionDicelossbatchnormalization
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

The paper is trying to establish that out-of-distribution (OoD) data does not have to come from an external source or carry extra annotations to help a fully-supervised biomedical segmentation network. It proposes Med-OoD, which builds a negative-supervision set from the existing in-distribution training images by erasing all foreground-class objects, prunes redundant OoD patches, and then estimates how much of that set to add so that the combined training set's positive-negative sample ratio approaches a reference value of 0.65. The claim is that this rebalancing suppresses the spurious foreground/background confusions that limited and imperfect medical datasets cause, improving mIoU and DSC for every architecture tested on the Lizard benchmark. A further claim is that a network trained entirely on OoD data, with no foreground labels at all, reaches 76.1% mIoU on the Lizard test set, which the authors offer as a new label-efficient learning paradigm. A sympathetic reader would care because the method changes no architecture and requires no external data, making the alleged gains cheap to test on any medical segmentation task.

What carries the argument

The load-bearing object is the OoD dataset constructed from the ID training set: foreground-class pixels are removed to make negative-only patches, and the unfiltered set is ranked by mIoU against an all-zero ground-truth mask so that patches the network already classifies correctly (mIoU = 1.0) are pruned out. The amount of OoD data is selected by a numerical estimator, Eq. (2)-(3), which searches over percentages pct to minimize |pnr(combined) - pnr_opt|, with pnr_opt = 0.65 taken as the optimal balance point from the ablation in Figure 5. Training then applies Dice loss against the normal class masks for ID images and against zero masks for OoD images, Eq. (4), with the two kinds of samples shuffled into the same batches so the network receives balanced positive and negative gradient signal. The role of this machinery is to convert a data-centric rebalancing operation into a concrete, architecture-agnostic training procedure.

What would settle it

Run Med-OoD on a second medical segmentation dataset with different foreground classes, keeping pnr_opt = 0.65 fixed; if the method does not beat the baseline or the ratio that maximizes validation mIoU falls outside (0.61, 0.68), the reference value is not a general constant. A direct check would be to compute the optimal pnr from a held-out validation split instead of the test set and compare it with 0.65.

Watch

Extended reading notes

Core claim

The paper's central discovery is that the balance between positive and negative pixel regions in the training data, rather than the sheer volume of data, drives a large share of foreground/background misclassification in medical segmentation. Starting from the Lizard training set, the authors remove the six foreground classes to create an unfiltered OoD set, then prune it by keeping only patches for which a baseline network predicts mIoU below 1.0, i.e., patches it mistakes as containing foreground; these are the negative samples that carry useful supervision. They then choose the OoD percentage that minimizes the absolute distance between the combined dataset's positive-negative ratio and pnr_opt = 0.65. In experiments, adding the selected OoD fraction improves mean IoU and Dice for Unet, Unet++, and MAnet variants, with the largest per-class gains on connective tissue (+19.6 mIoU), epithelial (+4.0), and lymphocyte (+2.8), while rare classes such as plasma, neutrophil, and eosinophil are unchanged. Training with only the selected OoD set, using zero masks as ground truth, yields 76.1% mIoU on the test set even though no foreground-class label is ever seen.

Load-bearing premise

The optimal positive-negative sample ratio of 0.65, which the method uses to choose how much OoD data to add, is taken as a stable reference point even though it was read off from ablation experiments on the same Lizard test data.

Editorial extensions

If this is right

  • If Med-OoD is right, any fully-supervised segmentation network can be improved on a limited medical dataset simply by reusing its own training images as negative samples, with no external data and no new annotations.
  • Adding more OoD data is not monotonically helpful: beyond the estimated optimal percentage, performance degrades below the baseline, so the fraction of OoD data must be chosen rather than maximized.
  • Batch normalization is part of the mechanism: removing BN layers cuts the Med-OoD gain, so the method's benefit depends on the normalization behavior of the target architecture.
  • The 76.1% mIoU OoD-only result implies that a substantial portion of segmentation structure can be learned from negative supervision alone, which could reduce annotation cost for foreground objects in some tasks.

Reading between the lines

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

  • Beyond the paper, the strongest testable extension is to check whether pnr_opt = 0.65 transfers to other medical domains; if it does, the method becomes a drop-in data recipe, and if not, the reported gains are partly a Lizard-specific tuning effect.
  • The OoD-only result suggests a broader label-efficiency principle: negative-only data can teach a network where object boundaries are not, which may generalize to instance segmentation or detection where foreground labels are the expensive part, though the paper only demonstrates nuclei segmentation.
  • The observation that more OoD data hurts past an optimum implies that data-centric methods should treat data quantity as a tunable hyperparameter rather than a monotonic resource; this is an editorial inference, not a claim the paper proves.
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 / 4 minor

Summary. The paper proposes Med-OoD, a data-centric framework for fully-supervised biomedical segmentation. It constructs an out-of-distribution (OoD) dataset by removing all foreground-class objects from the Lizard training images, prunes redundant samples, estimates an optimal percentage of OoD data by matching a positive-negative sample ratio to a reference pnr_opt=0.65, and trains with a Dice loss that supervises OoD samples with zero masks. Experiments on Lizard report consistent mIoU/DSC improvements across Unet, Unet++, and MAnet variants, and a claimed 'emerging paradigm' where training only on OoD data reaches 76.1% mIoU.

Significance. The problem addressed—mitigating foreground/background misclassification in medical segmentation without external data or extra annotations—is relevant, and the paper provides code and evaluates several architectures. The core idea of mining OoD samples from the task dataset itself is interesting. However, the central empirical claims are compromised by (i) an internal inconsistency in the OoD-only training result, and (ii) test-set-dependent selection of the key reference parameter pnr_opt. As a result, the paper does not currently establish a reliable method or benchmark.

major comments (3)
  1. [Sec. 3.1, Sec. 3.3 (Eq. 4), Table 4, Sec. 4.3 'Effectiveness of learning OoD data'] The reported 76.1% mIoU for training only on D_ood^opt is internally inconsistent with the described loss and data construction. Section 3.1 constructs D_ood by removing all foreground objects from Lizard, and Table 4 confirms that D_ood^opt contains zero samples for all six foreground classes. Equation (4) then trains OoD samples with the loss L_Dice(F_c(I_ood), 0). With an all-zero target, the Dice loss is minimized by predicting no foreground pixels for every class, so a network trained exclusively on D_ood^opt should produce all-background predictions and approximately 0% foreground mIoU on the test set. The paper instead reports 76.1% mIoU and, in Figure 6, per-class mIoUs identical to the baseline for several classes, which is impossible under the stated loss. This discrepancy indicates a mismatch among the data construction, the loss, or the evaluation metric (e.g., background included in mIoU). Because the abstract and contribution 5 present this as an 'emerging learning paradigm,' the claim is load-bearing and requires correction or verification via code.
  2. [Sec. 4.3, Figure 5, Eq. (2)-(3)] The optimal balance reference pnr_opt = 0.65 is derived from test-set performance on the same Lizard benchmark. Section 4.3 states that pnr_opt is set to 0.65 because the mIoU peak in Figure 5 occurs at pnr = 0.644, and Eq. (3) then selects the OoD percentage by minimizing the absolute difference to this reference. This makes the numerical estimation a post-hoc fit to the evaluation set: the reported test improvements of Med-OoD are partly a consequence of tuning a hyperparameter on the test data. The paper provides no validation split or sensitivity analysis to show that pnr_opt = 0.65 is stable across folds, architectures, or datasets. Without such evidence, the claim that Eqs. (2)-(3) 'rationally estimate' the OoD percentage is not supported.
  3. [Table 1, Sec. 4.1] The quantitative comparisons lack error bars or statistical tests. Each number in Table 1 is a single average over three folds, with no standard deviation or per-fold results, despite the 3-fold cross-validation protocol described in Sec. 4.1. The reported improvements are small (mIoU gains of 0.1 to 1.6 percentage points), and all experiments are on a single dataset. With no measure of variance and no independent validation of the pnr_opt choice, the claim that Med-OoD 'improves the results of all the methods' is not established.
minor comments (4)
  1. [Sec. 3.2] The definition of a positive/negative sample is ambiguous: an ID image patch containing both foreground and background pixels is counted as both positive and negative, so |D_pos| and |D_neg| are not disjoint and the pnr ratio's interpretation is unclear.
  2. [Sec. 3.1] The term 'unfiltered OoD set' is used both for unlabeled byproducts of the labeling process and for foreground-removed images from the ID set; this dual usage should be clarified to avoid confusion.
  3. [Figure 5] The x-axis of Figure 5 is described as pct decreasing from 100% to 40%, but the figure itself is not shown in the text; please make the axis labels and the relationship between pct and pnr explicit in the caption.
  4. [Throughout] Several typographical and reference errors appear (e.g., 'MAnet uses efficient MobileOnes' with reference ordering, 'Open-sampling' spelled inconsistently); a careful proofreading pass is recommended.

Circularity Check

1 steps flagged · score 7.0 of 10

The OoD proportion used for all reported gains is chosen by minimizing distance to pnr_opt=0.65, a value set from the test-set mIoU curve on the same Lizard benchmark; the reported improvements are therefore at least partly a test-set fit rather than an independent prediction.

  1. fitted input called prediction [Section 4.3 (Ablation study, Figure 5), operationalized in Section 3.2 Eqs. (2)-(3) and Section 4.1; evaluated in Table 1 on the same Lizard test set.]
    "pnr is increased from 0.527 to 0.725 as pct is decreased from 100% to 40% with an interval of 10%. It is seen that the segmentation performance attains the mIoU peak when pnr is 0.644 and the optimal value of pnr is expected to exist in the range of (0.61,0.68). Therefore, in all the experiments, we approximately set the optimal balance reference 𝑝𝑛𝑟𝑜𝑝𝑡 of the numerical estimation method to 0.65."

    Eq. (3) defines the OoD percentage via f_T(pct)=|pnr(D_com(pct*D_ood,D_id))-pnr_opt|, and Eq. (2) selects pct_opt as the minimizer. The reference pnr_opt=0.65 is not a first-principles constant; it is read off the Lizard test-set mIoU curve in Figure 5, where the mIoU peak occurs at pnr=0.644. All Table 1 test-set improvements are then produced by training on the combined dataset whose pnr is pulled toward this test-derived optimum and evaluated on that same Lizard test set. Thus the key selection step is an inverse lookup of the test-set-optimal ratio, and the claimed gains are confounded with hyperparameter fitting on the evaluation set. The paper provides no held-out or external validation to show that pnr_opt=0.65 is a stable, dataset-independent balance point.

full rationale

The central circularity is the test-set-derived optimal balance reference. Med-OoD's only free design parameter is the OoD fraction pct_opt, chosen by Eqs. (2)-(3) to make the training set's positive-negative ratio as close as possible to pnr_opt. The paper sets pnr_opt=0.65 because the Figure 5 ablation on the Lizard test set shows the mIoU peak at pnr=0.644. All main results (Table 1) are then reported on that same Lizard test set, so the reported improvements are at least partly a test-set fit rather than an independent verification of a general principle. The numerical 'estimation' method is therefore not a derivation from first principles; it is a search for the OoD amount that reproduces the empirically optimal pnr observed on the evaluation benchmark. There are no self-citations or imported uniqueness theorems, so this is a fitted-input circularity rather than a self-citation chain. Separately, the OoD-only 76.1% mIoU claim is difficult to reconcile with Eq. (4): if all OoD images are trained with Dice loss against zero foreground masks, the loss-minimizing prediction is no foreground, so one would expect near-zero foreground mIoU rather than 76.1%. That is an internal-consistency concern rather than a circularity, but it further weakens the 'emerging learning paradigm' claim. Taking both issues together, the paper's strongest headline results are not independently established, but the circularity itself is confined to the test-set-tuned pnr_opt selection, which warrants a 7 rather than a higher score.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The only free parameters are the OOD balance reference, the loss weight, and the pruning threshold, all of which are hand-set or test-set-tuned. The method introduces no new physical entities. The load-bearing assumptions are that label-based removal produces useful OOD data and that zero-mask Dice loss is an appropriate negative supervision signal.

free parameters (3)
  • pnr_opt = 0.65
    Optimal positive-negative sample ratio reference; set from test-set ablation (Figure 5) at peak pnr 0.644 and used in Eq. 3 to select OOD percentage.
  • lambda = 1.0
    Weight for the OOD Dice loss term in Eq. 4; no ablation is provided to justify this choice.
  • pruning threshold = mIoU = 1.0
    Hard threshold to keep only OOD samples that the ID-trained network predicts as pure background; hand-chosen without sensitivity analysis.
assumptions (3)
  • domain assumption Pixel-level annotations in Lizard are exact and can be used to mask out foreground objects to form valid background-only OOD patches.
    Section 3.1 constructs the unfiltered OOD set by removing foreground objects using labels; any annotation errors propagate into the OOD set and the zero-mask supervision.
  • domain assumption Dice loss against an all-zero target is a correct supervision signal for negative samples.
    Section 3.3, Eq. 4 applies standard Dice loss to zero masks; this assumes pushing predictions to zero for OOD patches helps generalization rather than merely suppressing foreground.
  • ad hoc to paper The Lizard test set is a valid within-distribution evaluation, and tuning pnr_opt on it does not invalidate the reported gains.
    Figure 5 uses test performance to set pnr_opt, yet the same benchmark is used to report final improvements.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Out-of-distribution data supervision towards biomedical semantic segmentation." pith.science (2026). https://pith.science/paper/KBXKHPE2

@misc{pith2026250712105,
  author       = {Pith},
  title        = {Pith review of: Out-of-distribution data supervision towards biomedical semantic segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KBXKHPE2}},
  note         = {Machine review of arXiv:2507.12105}
}
read the original abstract

Biomedical segmentation networks easily suffer from the unexpected misclassification between foreground and background objects when learning on limited and imperfect medical datasets. Inspired by the strong power of Out-of-Distribution (OoD) data on other visual tasks, we propose a data-centric framework, Med-OoD to address this issue by introducing OoD data supervision into fully-supervised biomedical segmentation with none of the following needs: (i) external data sources, (ii) feature regularization objectives, (iii) additional annotations. Our method can be seamlessly integrated into segmentation networks without any modification on the architectures. Extensive experiments show that Med-OoD largely prevents various segmentation networks from the pixel misclassification on medical images and achieves considerable performance improvements on Lizard dataset. We also present an emerging learning paradigm of training a medical segmentation network completely using OoD data devoid of foreground class labels, surprisingly turning out 76.1% mIoU as test result. We hope this learning paradigm will attract people to rethink the roles of OoD data. Code is made available at https://github.com/StudioYG/Med-OoD.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

62 extracted references · 54 canonical work pages

  1. [1]

    Hubmap-hacking the kidney,

    Adnane, D. and Tigoudern, A., “Hubmap-hacking the kidney,” Kaggle (2021)

    work page 2021

  2. [2]

    Revisiting neural scaling laws in language and vision,

    Alabdulmohsin, I., Neyshabur, B. and Zhai, X., “Revisiting neural scaling laws in language and vision,” Advances in Neural Information Processing Systems (2022)

    work page 2022

  3. [3]

    Deep semantic segmentation of natural and medical images: a review,

    Asgari Taghanaki, S., Abhishek, K., Cohen, J.P., Cohen -Adad, J. and Hamarneh, G. , “ Deep semantic segmentation of natural and medical images: a review,” Artificial Intelligence Review 54, 137–178 (2021)

    work page 2021

  4. [4]

    Out-distribution aware Self-training in an Open World Setting

    Augustin, M. and Hein, M. , “Out-distribution aware self -training in an open world setting,” arXiv preprint arXiv:2012.12372 (2020)

    work page Pith review arXiv 2020

  5. [5]

    Medical image segmentation via unsupervised convolutional neural network,

    Chen, J. and Frey, E.C., “Medical image segmentation via unsupervised convolutional neural network,” Medical Imaging with Deep Learning (2020)

    work page 2020

  6. [6]

    The Value of Out-of-Distribution Data

    De Silva, A., Ramesh, R., Priebe, C.E., Chaudhari, P. and Vogelstein, J.T., “The value of out-of-distribution data,” arXiv preprint arXiv:2208.10967 (2022)

    work page Pith review arXiv 2022

  7. [7]

    Imagenet: A large scale hierarchical image database,

    Deng, J., Dong, W., Socher, R., Li, L.J., Li, K. and Fei-Fei, L., “Imagenet: A large scale hierarchical image database,” IEEE conference on computer vision and pattern recognition, 248–255, Ieee (2009)

    work page 2009

  8. [8]

    Reducing network agnostophobia,

    Dhamija, A.R., Gü nther, M. and Boult, T., “Reducing network agnostophobia,” Advances in Neural Information Processing Systems (2018)

    work page 2018

Show all 62 references
  1. [9]

    Ma-net: A multi -scale attention network for liver and tumor segmentation,

    Fan, T., Wang, G., Li, Y. and Wang, H. , “Ma-net: A multi -scale attention network for liver and tumor segmentation,” IEEE Access 8, 179656–179665 (2020)

    2020

  2. [10]

    Improving the segmentation of anatomical structures in chest radiographs using u -net with an imagenet pre -trained encoder,

    Frid-Adar, M., Ben -Cohen, A., Amer, R. and Greenspan, H. , “Improving the segmentation of anatomical structures in chest radiographs using u -net with an imagenet pre -trained encoder,” Image Analysis for Moving Organ, Breast, and Thoracic Images: Third International Workshop...

    2018

  3. [11]

    Pannuke: an open pan -cancer histology dataset for nuclei instance segmentation and classification ,

    Gamper, J., Alemi Koohbanani , N., Benet, K., Khuram, A. and Rajpoot, N. , “Pannuke: an open pan -cancer histology dataset for nuclei instance segmentation and classification ,” Digital Pathology: 15th European Congress, ECDP 2019, Warwick, UK, April 10–13, 2019, Proceedings 1...

    2019

  4. [12]

    Are vision transformers robust to spurious correlations?

    Ghosal, S.S., Ming, Y. and Li, Y. , “ Are vision transformers robust to spurious correlations? ” ICML 2022: Workshop on Spurious Correlations, Invariance and Stability

    2022

  5. [13]

    Lizard: A large -scale dataset for colonic nuclear instance segmentation and classification ,

    Graham, S., Jahanifar, M., Azam, A., Nimir, M., Tsang, Y.W., Dodd, K., Hero, E., Sahota, H., Tank, A., Benes, K. and et al. , “Lizard: A large -scale dataset for colonic nuclear instance segmentation and classification ,” Proceedings of the IEEE/CVF International Conference on...

    2021

  6. [14]

    Hover-net: Simultaneous segmentation and classification of nuclei in multitissue histology images,

    Graham, S., Vu, Q.D., Raza, S.E.A., Azam, A., Tsang, Y.W., Kwak, J.T. and Rajpoot, N. , “Hover-net: Simultaneous segmentation and classification of nuclei in multitissue histology images,” Medical Image Analysis 58, 101563 (2019)

    2019

  7. [15]

    Unsupervised microvascular image segmentation using an active contours mimicking neural network,

    Gur, S., Wolf, L., Golgher, L. and Blinder, P., “Unsupervised microvascular image segmentation using an active contours mimicking neural network,” Proceedings of the IEEE/CVF international conference on computer vision, 10722–10731 (2019)

    2019

  8. [16]

    Deep convolutional neural networks for segmenting 3d in vivo multiphoton images of vasculature in alzheimer disease mouse models,

    Haft-Javaherian, M., Fang, L., Muse, V., Schaffer, C.B., Nishimura, N. and Sabuncu, M.R., “Deep convolutional neural networks for segmenting 3d in vivo multiphoton images of vasculature in alzheimer disease mouse models,” PloS one 14(3), e0213539 (2019)

    2019

  9. [17]

    Model rubik’s cube: Twisting resolution, depth and width for tinynets,

    Han, K., Wang, Y., Zhang, Q., Zhang, W., Xu, C. and Zhang, T., “Model rubik’s cube: Twisting resolution, depth and width for tinynets,” Advances in Neural Information Processing Systems 33, 19353–19364 (2020)

    2020

  10. [18]

    Deep residual learning for image recognition,

    He, K., Zhang, X., Ren, S. and Sun, J., “Deep residual learning for image recognition,” Proceedings of the IEEE conference on computer vision and pattern recognition, 770–778 (2016)

    2016

  11. [19]

    A baseline for detecting misclassified and out-of-distribution examples in neural networks,

    Hendrycks, D. and Gimpel, K., “A baseline for detecting misclassified and out-of-distribution examples in neural networks,” arXiv preprint arXiv:1610.02136 (2016)

    2016 arXiv

  12. [20]

    Deep anomaly detection with outlier exposure ,

    Hendrycks, D., Mazeika, M. and Dietterich, T., “Deep anomaly detection with outlier exposure ,” arXiv preprint arXiv:1812.04606 (2018)

    2018 arXiv

  13. [21]

    Generalized odin: Detecting out-of-distribution image without learning from out -of-distribution data ,

    Hsu, Y.C., Shen, Y., Jin, H. and Kira, Z., “Generalized odin: Detecting out-of-distribution image without learning from out -of-distribution data ,” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 10951–10960 (2020)

    2020

  14. [22]

    Densely connected convolutional networks ,

    Huang, G., Liu, Z., Van Der Maaten, L. and Weinberger, K.Q., “Densely connected convolutional networks ,” Proceedings of the IEEE conference on computer vision and pattern recognition , 4700–4708 (2017)

    2017

  15. [23]

    Ternausnet: U -net with vgg11 encoder pre -trained on imagenet for image segmentation,

    Iglovikov, V. and Shvets, A. , “Ternausnet: U -net with vgg11 encoder pre -trained on imagenet for image segmentation,” arXiv preprint arXiv:1801.05746 (2018)

    2018 arXiv

  16. [24]

    Batch normalization: Accelerating deep network training by reducing internal covariate shift,

    Ioffe, S. and Szegedy, C. , “Batch normalization: Accelerating deep network training by reducing internal covariate shift,” International conference on machine learning, 448–456, pmlr (2015)

    2015

  17. [25]

    Data-centric artificial intelligence,

    Jakubik, J., Vö ssing, M., Kü hl, N., Walk, J. and Satzger, G., “Data-centric artificial intelligence,” arXiv preprint arXiv:2212.11854 (2022)

    2022 arXiv

  18. [26]

    Scaling laws for neural language models,

    Kaplan, J., McCandlish, S., Henighan, T., Brown, T.B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J. and Amodei, D., “Scaling laws for neural language models,” arXiv preprint arXiv:2001.08361 (2020)

    2020 arXiv

  19. [27]

    Scaling laws for deep learning based image reconstruction ,

    Klug, T. and Heckel, R. , “Scaling laws for deep learning based image reconstruction ,” arXiv preprint arXiv:2209.13435 (2022)

    2022 arXiv

  20. [28]

    A multi-organ nucleus segmentation challenge,

    Kumar, N., Verma, R., Anand, D., Zhou, Y., Onder, O.F., Tsougenis, E., Chen, H., Heng, P.A., Li, J., Hu, Z. and et al., “A multi-organ nucleus segmentation challenge,” IEEE transactions on medical imaging 39(5), 1380–1391 (2019)

    2019

  21. [29]

    Weakly supervised semantic segmentation using out- of-distribution data,

    Lee, J., Oh, S.J., Yun, S., Choe, J., Kim, E. and Yoon, S., “Weakly supervised semantic segmentation using out- of-distribution data,” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 16897–16906 (2022)

    2022

  22. [30]

    Removing undesirable feature contributions using out-of- distribution data,

    Lee, S., Park, C., Lee, H., Yi, J., Lee, J. and Yoon, S., “Removing undesirable feature contributions using out-of- distribution data,” International Conference on Learning Representations

  23. [31]

    Analyzing overfitting under class imbalance in neural networks for image segmentation,

    Li, Z., Kamnitsas, K. and Glocker, B., “Analyzing overfitting under class imbalance in neural networks for image segmentation,” IEEE transactions on medical imaging 40(3), 1065–1077 (2020)

    2020

  24. [32]

    Cam-unet: Class activation map guided unet with feedback refinement for defect segmentation,

    Lin, D., Li, Y., Prasad, S., Nwe, T.L., Dong, S. and Oo, Z.M., “Cam-unet: Class activation map guided unet with feedback refinement for defect segmentation,” 2020 IEEE International Conference on Image Processing (ICIP), 2131–2135, IEEE (2020)

    2020

  25. [33]

    Microsoft coco: Common objects in context ,

    Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollá r, P. and Zitnick, C.L., “Microsoft coco: Common objects in context ,” Computer Vision -ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, 740–755....

    2014

  26. [34]

    V-net: Fully convolutional neural networks for volumetric medical image segmentation,

    Milletari, F., Navab, N. and Ahmadi, S.A., “V-net: Fully convolutional neural networks for volumetric medical image segmentation,” 2016 fourth international conference on 3D vision (3DV), 565–571, Ieee (2016)

    2016

  27. [35]

    On the impact of spurious correlation for out -of-distribution detection ,

    Ming, Y., Yin, H. and Li, Y. , “On the impact of spurious correlation for out -of-distribution detection ,” Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, 10051 –10059 (2022)

    2022

  28. [36]

    Bridging the gap between natural and medical images through deep colorization,

    Morra, L., Piano, L., Lamberti, F. and Tommasi, T. , “Bridging the gap between natural and medical images through deep colorization,” 2020 25th International Conference on Pattern Recognition (ICPR) , 835–842, IEEE (2021)

    2021

  29. [37]

    Towards a guideline for evaluation metrics in medical image segmentation,

    Mü ller, D., Soto-Rey, I. and Kramer, F. , “Towards a guideline for evaluation metrics in medical image segmentation,” BMC Research Notes 15(1) (2022)

    2022

  30. [38]

    Universal lesion detection and classification using limited data and weakly -supervised self -training,

    Naga, V., Mathai, T.S., Paul, A. and Summers, R.M., “Universal lesion detection and classification using limited data and weakly -supervised self -training,” Medical Image Learning with Limited and Noisy Data: First International Workshop, MIL -LanD 2022, Held in Conjunction w...

    2022

  31. [39]

    Medical image segmentation with limited supervision: A review of deep network models,

    Peng, J. and Wang, Y., “Medical image segmentation with limited supervision: A review of deep network models,” IEEE Access 9, 36827–36851 (2021)

    2021

  32. [40]

    Scaling laws for the few -shot adaptation of pre - trained image classifiers,

    Prato, G., Guiroy, S., Caballero, E., Rish, I. and Chandar, S., “Scaling laws for the few -shot adaptation of pre - trained image classifiers,” arXiv preprint arXiv:2110.06990 (2021)

    2021 arXiv

  33. [41]

    Nuisances via negativa: Adjusting for spurious correlations via data augmentation,

    Puli, A., Joshi, N., He, H. and Ranganath, R., “Nuisances via negativa: Adjusting for spurious correlations via data augmentation,” arXiv preprint arXiv:2210.01302 (2022)

    2022 arXiv

  34. [42]

    Generalized intersection over union: A metric and a loss for bounding box regression ,

    Rezatofighi, H., Tsoi, N., Gwak, J., Sadeghian, A., Reid, I. and Savarese, S., “Generalized intersection over union: A metric and a loss for bounding box regression ,” Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 658–666 (2019)

    2019

  35. [43]

    U-net: Convolutional networks for biomedical image segmentation,

    Ronneberger, O., Fischer, P. and Brox, T., “U-net: Convolutional networks for biomedical image segmentation,” Medical Image Computing and Computer -Assisted Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, 2...

    2015

  36. [44]

    The elephant in the room,

    Rosenfeld, A., Zemel, R. and Tsotsos, J.K., “The elephant in the room,” arXiv preprint arXiv:1808.03305 (2018)

    2018 arXiv

  37. [45]

    Mobilenetv2: Inverted residuals and linear bottlenecks,

    Sandler, M., Howard, A., Zhu, M., Zhmoginov, A. and Chen, L.C., “Mobilenetv2: Inverted residuals and linear bottlenecks,” Proceedings of the IEEE conference on computer vision and pattern recognition, 4510–4520 (2018)

    2018

  38. [46]

    How reliable are out -of-distribution generalization methods for medical image segmentation?

    Sanner, A., Gonzalez, C. and Mukhopadhyay, A., “How reliable are out -of-distribution generalization methods for medical image segmentation? ” Pattern Recognition: 43rd DAGMGermanConference, DAGM GCPR2021, Bonn, Germany, September 28–October 1, 2021, Proceedings, 604–617, Spri...

    2022

  39. [47]

    Global healthcare fairness: We should be sharing more, not less, data,

    Seastedt, K.P., Schwab, P., O’Brien, Z., Wakida, E., Herrera, K., Marcelo, P.G.F., Agha -Mir-Salim, L., Frigola, X.B., Ndulue, E.B., Marcelo, A. and et al., “Global healthcare fairness: We should be sharing more, not less, data,” PLOS Digital Health 1(10), e0000102 (2022)

    2022

  40. [48]

    Data-suite: Data -centric identification of in -distribution incongruous examples,

    Seedat, N., Crabbé , J. and van der Schaar, M. , “Data-suite: Data -centric identification of in -distribution incongruous examples,” International Conference on Machine Learning, 19467–19496, PMLR (2022)

    2022

  41. [49]

    Very deep convolutional networks for large -scale image recognition ,

    Simonyan, K. and Zisserman, A. , “Very deep convolutional networks for large -scale image recognition ,” International Conference on Learning Representations (2015)

    2015

  42. [50]

    Robustness to spurious correlations via human annotations ,

    Srivastava, M., Hashimoto, T. and Liang, P. , “Robustness to spurious correlations via human annotations ,” International Conference on Machine Learning, 91099119, PMLR (2020)

    2020

  43. [51]

    Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations ,

    Sudre, C.H., Li, W., Vercauteren, T., Ourselin, S. and Jorge Cardoso, M. , “Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations ,” Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Third Int...

    2017

  44. [52]

    Metrics for evaluating 3d medical image segmentation: analysis, selection, and tool,

    Taha, A.A. and Hanbury, A., “Metrics for evaluating 3d medical image segmentation: analysis, selection, and tool,” BMC medical imaging 15(1), 1–28 (2015)

    2015

  45. [53]

    Surrogate supervision for medical image analysis: Effective deep learning from limited quantities of labeled data ,

    Tajbakhsh, N., Hu, Y., Cao, J., Yan, X., Xiao, Y., Lu, Y., Liang, J., Terzopoulos, D. and Ding, X., “Surrogate supervision for medical image analysis: Effective deep learning from limited quantities of labeled data ,” 2019 IEEE 16th international symposium on biomedical imagin...

    2019

  46. [54]

    Embracing imperfect datasets: A review of deep learning solutions for medical image segmentation,

    Tajbakhsh, N., Jeyaseelan, L., Li, Q., Chiang, J.N., Wu, Z. and Ding, X., “Embracing imperfect datasets: A review of deep learning solutions for medical image segmentation,” Medical Image Analysis 63, 101693 (2020)

    2020

  47. [55]

    Deep learning convolutional networks for multiphoton microscopy vasculature segmentation,

    Teikari, P., Santos, M., Poon, C. and Hynynen, K. , “Deep learning convolutional networks for multiphoton microscopy vasculature segmentation,” arXiv preprint arXiv:1606.02382 (2016)

    2016 arXiv

  48. [56]

    An improved one millisecond mobile backbone ,

    Vasu, P.K.A., Gabriel, J., Zhu, J., Tuzel, O. and Ranjan, A., “An improved one millisecond mobile backbone ,” arXiv preprint arXiv:2206.04040 (2022)

    2022 arXiv

  49. [57]

    Open-sampling: Exploring out -of-distribution data for re - balancing long-tailed datasets,

    Wei, H., Tao, L., Xie, R., Feng, L. and An, B. , “Open-sampling: Exploring out -of-distribution data for re - balancing long-tailed datasets,” International Conference on Machine Learning, 23615–23630, PMLR (2022)

    2022

  50. [58]

    Boosting dense long -tailed object detection from data -centric view,

    Xu, W., Cao, J., Fu, T., Yao, H. and Wang, Y., “Boosting dense long -tailed object detection from data -centric view,” Proceedings of the Asian Conference on Computer Vision, 333–348 (2022)

    2022

  51. [59]

    Understanding rare spurious correlations in neural networks,

    Yang, Y.Y., Chou, C.N. and Chaudhuri, K., “Understanding rare spurious correlations in neural networks,” ICML 2022: Workshop on Spurious Correlations, Invariance and Stability

    2022

  52. [60]

    Mine your own anatomy: Revisiting medical image segmentation with extremely limited labels,

    You, C., Dai, W., Liu, F., Su, H., Zhang, X., Staib, L. and Duncan, J.S., “Mine your own anatomy: Revisiting medical image segmentation with extremely limited labels,” arXiv preprint arXiv:2209.13476 (2022)

    2022 arXiv

  53. [61]

    Exploiting the potential of datasets: A datacentric approach for model robustness,

    Zhong, Y., Wu, L., Liu, X. and Jiang, J., “Exploiting the potential of datasets: A datacentric approach for model robustness,” arXiv preprint arXiv:2203.05323 (2022)

    2022 arXiv

  54. [62]

    Unet++: A nested u -net architecture for medical image segmentation,

    Zhou, Z., Rahman Siddiquee, M.M., Tajbakhsh, N. and Liang, J. , “Unet++: A nested u -net architecture for medical image segmentation,” Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: 4th International Workshop, DLMIA 2018, and 8th...

    2018

Pith tools

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