Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Pixel-wise Modulated Dice Loss for Medical Image Segmentation

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper proposes a pixel-wise modulated Dice loss that weights each pixel by its current prediction error, and reports it outperforms existing difficulty-imbalance losses on three medical segmentation benchmarks, with the largest gains…

desk verdict A simple, plausible Dice variant with a pixel-wise focal term; the idea is worth testing, but small margins and test-set hyperparameter selection weaken the 'outperforms' claim. read the letter →

arxiv 2506.15744 v1 pith:J7URTDQJ submitted 2025-06-17 eess.IV cs.CVcs.LG

Seyed Mohsen Hosseini This is my paper
classification eess.IVcs.CVcs.LG
keywords Dicelossdifficultyimbalanceclassmedicalimagesegmentationpixel-wiseweightingfocalboundarycompound
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

Class imbalance and difficulty imbalance both degrade medical segmentation, and most current fixes combine a re-weighted cross-entropy loss with Dice loss. The paper claims that one simple change to Dice loss handles both: multiply each pixel's contribution by a dynamically changing weight $|y_p - \hat{p}_p|^{\gamma}$, with gradients blocked through the weight. The weight is near zero for confident correct pixels and near one for uncertain or wrong pixels, so the loss concentrates on boundaries and small structures without ranking pixels or adding a CE term. On three benchmarks—polyp, cardiac MRI, and MS lesion segmentation—the paper reports the best mean Dice, mean IoU, and normalized surface distance among the compared difficulty-imbalance losses, with the largest gains around boundaries.

What carries the argument

The load-bearing object is the modulating term $w_p^c=|y_p^c-\hat{p}_{p}^{c}|^{\gamma_c}$, a 2D map the same size as the prediction whose value is small where the pixel is confidently correct and near one where the prediction is wrong or uncertain. Inserted into the numerator and denominator of the per-class Dice score, it effectively replaces each class area by its difficult subset, rebalancing the background contribution, and because it is rebuilt every forward pass it keeps the loss focused as the network improves. Treating $w_p^c$ as a fixed hyperparameter with no gradient flow is what preserves the Dice-like geometry and keeps the extra computation limited to elementwise operations.

What would settle it

Retrain the Kvasir-SEG experiment with identical settings except that gradients are allowed to flow through $w_p^c$; if the mean Dice and normalized surface distance do not fall below the detached version, the paper's central condition is not the cause of the improvement. A complementary check is whether, on a fixed set of hard pixels, the gradient of the detached PM Dice loss is aligned with the direction that improves the plain Dice score.

Watch

Extended reading notes

Core claim

The discovery is that inserting a per-pixel error-based weight inside the Dice formula turns a class-imbalance loss into a difficulty-aware one without changing its geometric character. The proposed loss is $$\mathrm{PM Dice}=1-\frac{1}{C}\sum_{c=1}^{C}\frac{2\sum_p w_p^c y_p^c \hat{p}_{p}^{c}+\epsilon}{\sum_p w_p^c [(y_p^c)^2+(\hat{p}_{p}^{c})^2]+\epsilon},$$ with $w_p^c=|y_p^c-\hat{p}_{p}^{c}|^{\gamma_c}$, where $\hat{p}_{p}^{c}$ is the network's predicted probability with no gradient update. The paper asserts that this stop-gradient condition is essential: without it the objective stops being a weighted Dice loss, so reducing it would not track the evaluation metric. Reported results on Kvasir-SEG, ACDC, and MSSEG-2 show PM Dice achieving the highest mean Dice, mean IoU, and normalized surface distance among the compared losses, and the paper concludes that combining it with cross-entropy is unnecessary.

Load-bearing premise

The load-bearing premise is that computing the pixel weight $|y_p^c-\hat{p}_{p}^{c}|^{\gamma}$ from the current prediction while allowing no gradient to flow through it is what focuses training on difficult pixels; the paper states this condition is essential and that performance drops without it, but it gives no derivation, ablation, or loss-surface analysis to support that claim.

Editorial extensions

If this is right

  • A U-Net trained with PM Dice can drop the cross-entropy term entirely, saving practitioners the effort of tuning a compound-loss weight.
  • Because the method avoids per-image error ranking, it adds only elementwise multiplications to the Dice computation and scales to large 3D volumes.
  • Per-class focusing parameters give an explicit control for severe imbalance, as used on MSSEG where foreground and background get different gamma values.
  • The consistently highest normalized surface distance values reported imply that boundary accuracy improves without any explicit edge-aware term.
  • PM Dice offers both pixel-level and class-level focusing, so it can replace Focal CE, TopK Dice, and compound-loss schemes with a single loss function.

Reading between the lines

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

  • The stop-gradient condition is asserted but not demonstrated; a controlled ablation that lets gradients flow through $w_p^c$ would show whether the detached weighting is the mechanism or merely a helpful regularizer.
  • Since $w_p^c$ is peaked at object boundaries and small structures, PM Dice may be acting as an implicit boundary-aware loss; comparing it against explicit boundary losses on the same tasks would separate boundary emphasis from difficulty weighting.
  • The difficulty signal is the raw prediction error on an uncalibrated softmax, so if probability calibration is poor the weight will misidentify hard pixels; calibration-aware variants could behave differently.
  • The paper defines a voxel-wise version but reports no 3D experiments, so applying VM Dice to volumetric segmentation would test whether the pixel-level advantage transfers to voxel grids.
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 manuscript proposes the Pixel-wise Modulated Dice (PM Dice) loss, defined in Eq. (5) as the standard multi-class Dice loss with each pixel's contribution reweighted by a detached modulating term w_p^c = |y_p^c - \hat p_p^c|^\gamma. The authors argue that this term focuses training on difficult pixels while preserving Dice's robustness to class imbalance, and they claim that the proposed loss outperforms existing difficulty-imbalance losses on three medical segmentation datasets: Kvasir-SEG, ACDC, and MSSEG-2. Experiments use a U-Net with identical training conditions across losses, and the paper reports mean Dice, IoU, NSD, precision, and recall over three runs. The central claims are that PM Dice improves boundary accuracy and removes the need for a compound CE loss.

Significance. If the claimed gains are real, the contribution is practically useful: the modification is simple, adds negligible computation, and avoids the ranking step of TopK-style losses. The choice of datasets covers binary and multi-class tasks with different imbalance levels, and the NSD results suggest a boundary-focused benefit. The paper also makes a falsifiable design claim about detaching the modulating term. However, the current evidence does not yet establish the central outperformance claim: the MSSEG hyperparameters are selected on the test split, and no variance estimates are reported for any table. The contribution is therefore promising but needs a strengthened evaluation protocol before the 'outperforms other methods' conclusion can be accepted.

major comments (3)
  1. [§3.1–3.2, Table 3] The MSSEG comparison is not a fair test of the proposed loss. Section 3.1 describes only a random train/test split (9 scans for training, the rest for testing) with no validation set, yet Section 3.2 and Figures 3–4 state that the foreground and background focusing parameters γ_fg=2 and γ_bg=1 were selected from sweeps on the same dataset. Since these sweeps are evaluated on the test split that later produces Table 3, the 0.63-point mDice advantage over TopK CE + TopK Dice can be explained by test-set hyperparameter selection. Please use a held-out validation split, or report results for all swept γ values on the test set, and restrict any 'best' claim to configurations chosen without test-set information.
  2. [§3.2, Tables 1–3] The paper reports each entry as the average of 3 independent runs but gives no standard deviation, confidence interval, or per-run values. Several decisive margins are very small: 0.04 mDice on Kvasir (Table 1, 91.88 vs 91.84), 0.50 mDice on ACDC (Table 2, 90.61 vs 90.11), and 0.63 mDice on MSSEG (Table 3, 69.07 vs 68.44). With three runs, these margins are within plausible run-to-run variability, so the central claim that PM Dice 'outperforms other methods' is not statistically supported. Please report standard deviations or individual run results and, where possible, a paired significance test across runs.
  3. [§2, Eq. (5)] The design rests on detaching the modulating term w_p^c = |y_p^c - \hat p_p^c|^γ from gradient computation, and the paper states that 'without this key condition the performance would drop' (Section 2). This assertion is not supported by any experiment, derivation, or loss-surface analysis in the manuscript. Because the detached weighting changes the objective being minimized, it is not obvious that minimizing Eq. (5) is more aligned with the Dice score than the version with gradient flow through w_p, or than plain Dice. Please add an ablation comparing the detached and non-detached variants, together with at least one alternative weighting, on at least one dataset; otherwise the central outperformance claim is tied to an unverified design decision.
minor comments (4)
  1. [§2, Eqs. (1)–(3)] Equation (1) appears to define plain cross-entropy, while Eq. (2) is labeled TopK CE, but the rendering makes the definitions of T, N, and the pixel sets hard to follow; please rewrite the equations and define every symbol in one place.
  2. [Fig. 2] The caption of Figure 2 lists focusing values '4.5, 7, 8, 5', which appears to be a typo for '0.5, 1, 2, 5'; also clarify how the loss values in that figure were computed (e.g., a single image at a fixed checkpoint).
  3. [§1 and §2] There are several typographical errors, including 'sate of the art' for 'state of the art' and 'variate' for 'variant' in Section 2; a careful proofread is recommended.
  4. [References] The reference list contains uncited entries (e.g., [7]–[11]) and duplicates the Kvasir-SEG dataset citation as [8] and [43]; please reconcile the bibliography with the in-text citations.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the PM Dice loss is defined independently of the reported results, and the claimed improvements are empirical comparisons against external benchmarks rather than consequences of the loss definition.

full rationale

The proposed loss in Eq. 5 is a direct modification of the standard Dice loss in Eq. 4: a pixel-wise modulating term w_p = |y_p - p_hat_p|^gamma multiplies both numerator and denominator, and p_hat is explicitly detached from the gradient computation. Nothing in Eq. 5 is defined in terms of the reported mDice, mIoU, or NSD tables, and the evaluation metrics are not computed from the loss itself. The central 'outperforms' claim is an empirical comparison on three external datasets under the same network, data, augmentation, and learning-rate settings, with the only varying factor being the loss function. The comparison includes the author's own TopK Dice [42] as a baseline, but that self-citation is not load-bearing: the numeric comparison would stand regardless of who authored TopK Dice. The unsupported assertion that detaching the modulating term is a 'key condition' is an optimization claim, not a circular definition, and the MSSEG gamma tuning on the test set (Figures 3-4 followed by Table 3) is a statistical-validity concern rather than a constructional circularity. No equation in the paper reduces to its own input, and no fitted parameter is renamed as a prediction. The derivation chain is therefore self-contained with respect to the reported evaluations.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim depends on a small number of hand-set hyperparameters (gamma per class, epsilon) and on three domain assumptions: Dice loss is a suitable base, the stop-gradient design is a valid training objective, and the shared U-Net setup is fair across all compared losses. No new physical or conceptual entities are introduced.

free parameters (2)
  • Focusing parameter gamma (per class) = gamma=1 for Kvasir and ACDC; gamma_fg=2, gamma_bg=1 for MSSEG
    Controls the exponent of the pixel-wise modulating term |y - p_hat|^gamma. Values were chosen per dataset, and for MSSEG selected from test-set curves in Figures 3-4, so they are hand-tuned rather than derived.
  • Smoothing factor epsilon = 1e-5
    Small constant added to numerator and denominator to avoid division by zero; chosen to be small so that easy-pixel losses do not collapse. Not load-bearing but hand-set.
assumptions (3)
  • domain assumption Dice loss is an effective base loss for class imbalance in medical segmentation
    The method assumes Dice loss handles class imbalance better than CE, citing prior surveys (refs [1,5]) without re-deriving this property in the proposed setting.
  • ad hoc to paper Detaching the modulating term (no gradient through w) preserves a useful training objective
    The paper states that without stop-gradient 'the performance would drop' (Section 2), but gives no analysis of the resulting loss surface or proof that minimizing this weighted loss tracks Dice score.
  • domain assumption The fixed U-Net training setup is equally suitable for all compared losses
    The comparison keeps architecture, augmentation, and schedule fixed across losses, implicitly assuming every loss is fairly tuned; the paper does not report per-loss hyperparameter optimization beyond the proposed gamma.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Pixel-wise Modulated Dice Loss for Medical Image Segmentation." pith.science (2026). https://pith.science/paper/J7URTDQJ

@misc{pith2026250615744,
  author       = {Pith},
  title        = {Pith review of: Pixel-wise Modulated Dice Loss for Medical Image Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J7URTDQJ}},
  note         = {Machine review of arXiv:2506.15744}
}
read the original abstract

Class imbalance and the difficulty imbalance are the two types of data imbalance that affect the performance of neural networks in medical segmentation tasks. In class imbalance the loss is dominated by the majority classes and in difficulty imbalance the loss is dominated by easy to classify pixels. This leads to an ineffective training. Dice loss, which is based on a geometrical metric, is very effective in addressing the class imbalance compared to the cross entropy (CE) loss, which is adopted directly from classification tasks. To address the difficulty imbalance, the common approach is employing a re-weighted CE loss or a modified Dice loss to focus the training on difficult to classify areas. The existing modification methods are computationally costly and with limited success. In this study we propose a simple modification to the Dice loss with minimal computational cost. With a pixel level modulating term, we take advantage of the effectiveness of Dice loss in handling the class imbalance to also handle the difficulty imbalance. Results on three commonly used medical segmentation tasks show that the proposed Pixel-wise Modulated Dice loss (PM Dice loss) outperforms other methods, which are designed to tackle the difficulty imbalance problem.

Figures

Figures reproduced from arXiv: 2506.15744 by the authors.

Figure 1
Figure 1. Pixel-wise Modulated (PM) Dice loss. In a [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. PM Dice loss for different focusing param [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The effect of changing the focusing param [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: The effect of changing the focusing param [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Visual comparison of results from Kvasir, [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. DAMamba-UNet3D: A Parameter-Efficient Mamba State Space U-Net with Dynamic Adaptive Scan for 3D Medical Image Segmentation

    cs.CV 2026-07 conditional novelty 6.0 of 10

    DAMamba-UNet3D combines encoder-only tri-plane Dynamic Adaptive Scan with a convolutional U-Net, reaching 0.815 mean Dice on BraTS 2020 at 5.3M parameters.

Reference graph

Works this paper leans on

47 extracted references · 37 canonical work pages · cited by 1 Pith paper

  1. [1]

    Loss odyssey in medical image segmentation

    Ma, Jun, et al. "Loss odyssey in medical image segmentation." Medical Image Analysis 71 (2021): 102035

    work page 2021

  2. [2]

    Focal loss for dense object detection

    Lin, Tsung-Yi, et al. "Focal loss for dense object detection." Proceedings of the IEEE international conference on computer vision. 2017

    work page 2017

  3. [3]

    Full-resolution residual ne tworks for semantic segmentation in street scenes

    Pohlen, Tobias, et al. "Full-resolution residual ne tworks for semantic segmentation in street scenes." Proceeding s of the IEEE conference on computer vision and pattern recognition. 2017

    work page 2017

  4. [4]

    U- net: Convolutional networks for biomedical image segmentation

    Ronneberger, Olaf, Philipp Fischer, and Thomas Brox . "U- net: Convolutional networks for biomedical image segmentation." Medical Image Computing and Computer - Assisted Intervention–MICCAI 2015: 18th Internation al Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. Springer International Pu blishing, 2015

    work page 2015

  5. [5]

    Optimizing the dice score and jaccard index for medical image segmentation: Theory and practice

    Bertels, Jeroen, et al. "Optimizing the dice score and jaccard index for medical image segmentation: Theory and practice." Medical Image Computing and Computer Assisted Intervention–MICCAI 2019: 22nd Internation al Conference, Shenzhen, China, October 13–17, 2019, Proceedings, Part II 22. Springer International Pub lishing, 2019

    work page 2019

  6. [6]

    Deep residual learning for ima ge recognition

    He, Kaiming, et al. "Deep residual learning for ima ge recognition." Proceedings of the IEEE conference on computer vision and pattern recognition. 2016

    work page 2016

  7. [7]

    FCB-Swin V2 Transformer for Polyp Segmentation

    Fitzgerald, Kerr, and Bogdan Matuszewski. "FCB-Swin V2 Transformer for Polyp Segmentation." arXiv preprint arXiv:2302.01027 (2023)

    arXiv 2023

  8. [8]

    Kvasir-SEG: A Segmented Polyp Dataset,

    D. Jha, P.H. Smedsrud, M.A. Riegler, P. Halvorsen, T. de Lange, D. Johansen & H.D. Johansen , "Kvasir-SEG: A Segmented Polyp Dataset," in International Conferen ce on Multimedia Modeling, 2019

    work page 2019

Show all 47 references
  1. [9]

    WM-DOVA maps for accurate polyp highlighting in colonoscopy: Validat ion vs. saliency maps from physicians,

    J. Bernal, F. J. Sanchez, G. Fern ´ andez-Esparrach , D. Gil, C. Rodríguez & F. Vilariño,, "WM-DOVA maps for accurate polyp highlighting in colonoscopy: Validat ion vs. saliency maps from physicians," Computerized Medica l Imaging and Graphics, vol. 43, pp. 99-111, 2015

    2015

  2. [10]

    Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic)

    Noel Codella, Veronica Rotemberg, Philipp Tschandl, M Emre Celebi, Stephen Dusza, David Gutman, Brian Hel ba, Aadi Kalloo, Konstantinos Liopyris, Michael Marchet ti, et al. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imagin...

    2018 arXiv

  3. [11]

    In: International Confere nce on Medical Image Computing and Computer-Assisted Intervention

    Fu, S., Lu, Y., Wang, Y., Zhou, Y., Shen, W., Fishm an, E., Yuille, A.: Domain adaptive relational reasoning fo r 3d multi-organ segmentation. In: International Confere nce on Medical Image Computing and Computer-Assisted Intervention. pp. 656-666. Springer (2020)

    2020

  4. [12]

    Transunet: Transformers make strong encoders for me dical image segmentation

    Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Eh san Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zh ou. Transunet: Transformers make strong encoders for me dical image segmentation. arXiv preprint arXiv:2102.04306 , 2021

    2021 arXiv

  5. [13]

    Pranet: Parallel reverse at tention network for polyp segmentation

    Fan, Deng-Ping, et al. "Pranet: Parallel reverse at tention network for polyp segmentation." International conf erence on medical image computing and computer-assisted intervention. Cham: Springer International Publishi ng, 2020

    2020

  6. [14]

    Uacanet: Uncertainty augmented context attention for polyp segmentation

    Kim, Taehun, Hyemin Lee, and Daijin Kim. "Uacanet: Uncertainty augmented context attention for polyp segmentation." Proceedings of the 29th ACM Internat ional Conference on Multimedia. 2021

    2021

  7. [15]

    nnU-Net for brain tumor segmentation

    Isensee, Fabian, et al. "nnU-Net for brain tumor segmentation." Brainlesion: Glioma, Multiple Sclero sis, Stroke and Traumatic Brain Injuries: 6th Internatio nal Workshop, BrainLes 2020, Held in Conjunction with MICCAI 2020, Lima, Peru, October 4, 2020, Revised Selected Papers...

    2020

  8. [16]

    Multi- scale Hierarchical Vision Transformer with Cascaded Attention Decoding for Medical Image Segmentation

    Rahman, Md Mostafijur, and Radu Marculescu. "Multi- scale Hierarchical Vision Transformer with Cascaded Attention Decoding for Medical Image Segmentation." arXiv preprint arXiv:2303.16892 (2023)

    2023 arXiv

  9. [17]

    Class-balanced loss based on eff ective number of samples

    Cui, Yin, et al. "Class-balanced loss based on eff ective number of samples." Proceedings of the IEEE/CVF conference on computer vision and pattern recogniti on. 2019

    2019

  10. [18]

    In Proceedings of the IEEE international conference on computer vision, p ages 1395–1403, 2015

    Holistically-nested edge detection. In Proceedings of the IEEE international conference on computer vision, p ages 1395–1403, 2015

    2015

  11. [19]

    Universal loss reweightin g to balance lesion size inequality in 3D medical image segmentation

    Shirokikh, Boris, et al. "Universal loss reweightin g to balance lesion size inequality in 3D medical image segmentation." Medical Image Computing and Computer Assisted Intervention–MICCAI 2020: 23rd Internation al Conference, Lima, Peru, October 4–8, 2020, Proceedi ngs, Part ...

    2020

  12. [20]

    Learning from imbalanced data

    He, Haibo, and Edwardo A. Garcia. "Learning from imbalanced data." IEEE Transactions on knowledge an d data engineering 21.9 (2009): 1263-1284

    2009

  13. [21]

    Rethinking atrous convolution for semantic image segmentation

    Chen, Liang-Chieh, et al. "Rethinking atrous convolution for semantic image segmentation." arXiv preprint arXiv:1706.05587 (2017)

    2017 arXiv

  14. [22]

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

    Sudre, Carole H., et al. "Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations." Deep Learning in Medical Image Analysis and Multimo dal Learning for Clinical Decision Support: Third Inter national 11 Workshop, DLMIA 2017, and 7th Inter...

    2017

  15. [23]

    Tversky loss function for image segment ation using 3D fully convolutional deep networks

    Salehi, Seyed Sadegh Mohseni, Deniz Erdogmus, and Ali Gholipour. "Tversky loss function for image segment ation using 3D fully convolutional deep networks." Intern ational workshop on machine learning in medical imaging. Ch am: Springer International Publishing, 2017

    2017

  16. [24]

    Combo loss: Handling input and output imbalance in multi-organ segmentation

    Taghanaki, Saeid Asgari, et al. "Combo loss: Handling input and output imbalance in multi-organ segmentation." Computerized Medical Imaging and Graphics 75 (2019): 24-33

    2019

  17. [25]

    Unet++: Redesigning skip connections to exploit multiscale features in image segmentation

    Zhou, Zongwei, et al. "Unet++: Redesigning skip connections to exploit multiscale features in image segmentation." IEEE transactions on medical imaging 39.6 (2019): 1856-1867

    2019

  18. [26]

    Class-wise difficulty-balanced loss for solving cl ass- imbalance

    Sinha, Saptarshi, Hiroki Ohashi, and Katsuyuki Naka mura. "Class-wise difficulty-balanced loss for solving cl ass- imbalance." Proceedings of the Asian conference on computer vision. 2020

    2020

  19. [27]

    Bridging category-level and instance-level semantic image segmentation

    Wu, Zifeng, Chunhua Shen, and Anton van den Hengel. "Bridging category-level and instance-level semantic image segmentation." arXiv preprint arXiv:1605.06885 (2016)

    2016 arXiv

  20. [28]

    Learning deep representation f or imbalanced classification

    Huang, Chen, et al. "Learning deep representation f or imbalanced classification." Proceedings of the IEEE conference on computer vision and pattern recogniti on. 2016

    2016

  21. [29]

    Automated volumetric assessment w ith artificial neural networks might enable a more accu rate assessment of disease burden in patients with multi ple sclerosis

    Brugnara, Gianluca, Fabian Isensee, Ulf Neuberger, David Bonekamp, Jens Petersen, Ricarda Diem, Brigitte Wildemann et al. "Automated volumetric assessment w ith artificial neural networks might enable a more accu rate assessment of disease burden in patients with multi ple scl...

    2020

  22. [30]

    Deep learning to achieve clinicall y applicable segmentation of head and neck anatomy fo r radiotherapy

    Stanislav Nikolov, Sam Blackwell, Alexei Zverovitch , Ruheena Mendes, Michelle Livne, Jeffrey De Fauw, Yo jan Patel, Clemens Meyer, Harry Askham, Bernardino Romera- Paredes, et al. “Deep learning to achieve clinicall y applicable segmentation of head and neck anatomy fo r radi...

    arXiv

  23. [31]

    Common limitations of image processing metrics: A picture story

    Reinke, Annika, et al. "Common limitations of image processing metrics: A picture story." arXiv preprin t arXiv:2104.05642 (2021)

    2021 arXiv

  24. [32]

    V-net: Fully convolutional neural networks for vol umetric medical image segmentation

    Milletari, Fausto, Nassir Navab, and Seyed-Ahmad Ahmadi. "V-net: Fully convolutional neural networks for vol umetric medical image segmentation." 2016 fourth internatio nal conference on 3D vision (3DV). Ieee, 2016

    2016

  25. [33]

    Balancing the Scales: A Comprehensive Study on Tackling Class Imbalance in Binary Classification

    Abdelhamid, Mohamed, and Abhyuday Desai. "Balancing the Scales: A Comprehensive Study on Tackling Class Imbalance in Binary Classification." arXiv preprint arXiv:2409.19751 (2024)

    2024 arXiv

  26. [34]

    SMOTE: synthetic minority over ‐ sampling technique

    Nitesh, V. Chawla. "SMOTE: synthetic minority over ‐ sampling technique." J Artif Intell Res 16.1 (2002)

    2002

  27. [35]

    One- class classification: A survey

    Perera, Pramuditha, Poojan Oza, and Vishal M. Patel. "One- class classification: A survey." arXiv preprint arXiv:2101.03064 (2021)

    2021 arXiv

  28. [36]

    Handling imbalanced medical image data: A deep-learning-based one-class classification approach

    Gao, Long, et al. "Handling imbalanced medical image data: A deep-learning-based one-class classification approach." Artificial intelligence in medicine 108 (2020)

    2020

  29. [37]

    Handling imbalanced medica l datasets: review of a decade of research

    Salmi, Mabrouka, et al. "Handling imbalanced medica l datasets: review of a decade of research." Artifici al Intelligence Review 57.10 (2024)

    2024

  30. [38]

    3D segmentation with exponent ial logarithmic loss for highly unbalanced object sizes

    Wong, Ken CL, et al. "3D segmentation with exponent ial logarithmic loss for highly unbalanced object sizes." Medical Image Computing and Computer Assist ed Intervention–MICCAI 2018: 21st International Confer ence, Granada, Spain, September 16-20, 2018, Proceedings, Part III 1...

    2018

  31. [39]

    Focal dice loss and image dilation for brain tumor segmentation

    Wang, Pei, and Albert CS Chung. "Focal dice loss and image dilation for brain tumor segmentation." Deep Learni ng in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: 4th International Worksh op, DLMIA 2018, and 8th International Workshop, ML-CDS 20...

    2018

  32. [40]

    Focal dice loss-based V-Net for liver segments classification

    Prencipe, Berardino, et al. "Focal dice loss-based V-Net for liver segments classification." Applied Sciences 12 .7 (2022)

    2022

  33. [41]

    Rethinking dice loss for me dical image segmentation

    Zhao, Rongjian, et al. "Rethinking dice loss for me dical image segmentation." 2020 IEEE international confer ence on data mining (ICDM). IEEE, 2020

    2020

  34. [42]

    topK dice loss for medical image segmentation

    Seyed mohsen Hosseini. “topK dice loss for medical image segmentation.” 35th British Machine Vision Conferen ce

  35. [43]

    In: Multimedia Modeling: 26th Internationa l Conference, MMM (2020)

    Jha D, Smedsrud PH, Riegler MA, Halvorsen P, de Lange T, Johansen D, Johansen HD.: Kvasir-seg: A segmented polyp dataset. In: Multimedia Modeling: 26th Internationa l Conference, MMM (2020)

    2020

  36. [44]

    Deep learning techniques for automatic MRI cardiac multi-structures segmentation and diagnosis: is the problem solved?

    Bernard, Olivier, et al. "Deep learning techniques for automatic MRI cardiac multi-structures segmentation and diagnosis: is the problem solved?." IEEE transactio ns on medical imaging 37.11 (2018)

    2018

  37. [45]

    MSSEG-2 challenge proceedings: Multiple sclerosis new lesions segment ation challenge using a data management and processing infrastructure

    Commowick, Olivier, et al. "MSSEG-2 challenge proceedings: Multiple sclerosis new lesions segment ation challenge using a data management and processing infrastructure." MICCAI 2021-24th international conference on medical image computing and computer assisted intervention. 2021

    2021

  38. [46]

    Double encoder-decoder networks for gastrointestinal polyp segmentation

    Galdran, Adrian, Gustavo Carneiro, and Miguel A. González Ballester. "Double encoder-decoder networks for gastrointestinal polyp segmentation." International Conference on Pattern Recognition. Cham: Springer International Publishing, 2021

    2021

  39. [2024]

    Glasgow, UK, November 25-28, 2024

    BMVC 2024. Glasgow, UK, November 25-28, 2024

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.