Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

MAMBO-NET: Multi-Causal Aware Modeling Backdoor-Intervention Optimization for Medical Image Segmentation Network

T0 review · 4 major / 5 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 Medical segmentation models, MAMBO-NET argues, are biased by hidden confounders; re-fitting them with a backdoor causal adjustment lifts Dice by up to 2.28% across five datasets.

desk verdict A real architecture with a broken causal story: Eq. (9) is not the backdoor adjustment it claims to be, but the multi-dataset results are worth a referee's time. read the letter →

arxiv 2505.21874 v1 pith:3NYX3KOR submitted 2025-05-28 eess.IV cs.CV

Ruiguo Yu , Yiyang Zhang , Yuan Tian , Yujie Diao , Di Jin , Witold Pedrycz This is my paper
classification eess.IVcs.CV
keywords medicalimagesegmentationcausalinferencebackdooradjustmentconfusionfactorsGaussianmixturemodelinguncertainty-awarelossultrasound
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

Standard segmentation networks fit the conditional distribution $P(Y'|X)$ and, the paper argues, thereby inherit bias from confusion factors — anatomy, imaging limitations, acquisition habits — that affect both the image and the network's output. MAMBO-NET instead learns a backdoor-adjusted distribution $P(Y'|\mathrm{do}(X))$ — the classical causal-intervention sum over confounder states — modeling the hidden confounders as $K$ trainable Gaussian distributions derived from the image and constrained by the ground-truth mask, then fusing them into every decoder stage through a learnable weighting matrix. The paper's empirical claim is that on three ultrasound datasets this raises Dice by up to 2.28% over its UNeXt backbone and cuts the false discovery rate on dermoscopy and colonoscopy datasets by 1.49% and 1.87%, respectively. If true, it gives medical segmentation a general deconfounding mechanism that needs no manual labeling of the confounding concepts.

What carries the argument

The load-bearing object is the backdoor-adjustment identity $P(Y'|\mathrm{do}(X)) = \sum_c P(Y'|X,c)P(c)$, implemented in a trainable, implicit form: $K$ Gaussian priors sampled by reparameterization stand in for the confounder set $C$, a posterior branch constrains them through KL divergence, and the CIBM module computes $\Omega \times Z$ as a per-channel weighted sum of sampled confounder features spliced into every decoder stage. This is the step that converts a correlation-fitted segmenter into one that claims to intervene on $X$ by conditioning on $C$, and it is also the step that carries all of the causal semantics.

What would settle it

Create a synthetic segmentation dataset with an injected, measurable confounder, such as an additive speckle artifact whose per-image strength is known. Train MAMBO-NET and check whether the learned Gaussian components track the injected confounder values, and whether removing the confounder from the test images erases the Dice gain over the backbone; if the intervention branch still changes predictions when the confounder is absent, the module is conditioning on image statistics rather than performing the claimed causal adjustment.

Watch

Extended reading notes

Core claim

The paper's central claim is that fitting $P(Y'|X)$ is biased because confusion factors $C$ — anatomical variation, imaging modality limits, acquisition habits — influence both the observed image and the network's prediction, and MAMBO-NET corrects this by learning $P(Y'|\mathrm{do}(X)) = \sum_c P(Y'|X,c)P(c)$. The confounders are not specified by hand: a Gaussian Self-Modeling branch fits $K$ prior distributions $N(\mu_p^{(i)}(X), \sigma_p^{(i)2}(X))$ from the image, a posterior branch fits the same number of distributions from the ground-truth mask, and a KL-divergence loss together with a boundary-focused uncertainty loss align them. The CIBM intervention branch then samples confounder features $Z$ and combines them into each decoder stage as $\Omega \times Z$, a learnable channel-wise weighting whose rows sum to one. On three ultrasound datasets the method is reported to beat its UNeXt backbone by up to 2.28% in Dice, and to reduce the false discovery rate by 1.49% on ISIC2018 and 1.87% on KVASIR, with ablations attributing the gain to the combination of the two new modules.

Load-bearing premise

If the learned weighted sum of Gaussian samples does not truly implement the backdoor adjustment it approximates — an equivalence the paper asserts without formal proof — the method becomes a feature-conditioning module and the claimed removal of confusion factors collapses.

Editorial extensions

If this is right

  • Predictions conditioned on $P(Y'|\mathrm{do}(X))$ should degrade less than the plain $P(Y'|X)$ baseline when the acquisition distribution shifts — a new scanner, operator, or protocol — because confounders are adjusted for rather than memorized.
  • The lower false-discovery rates translate into fewer false-positive lesion regions presented to clinicians in screening workloads, where each flagged region triggers a follow-up examination.
  • The ablation curve for the number of Gaussian components supplies a tuning rule: enough components to cover the dataset's latent confounders ($K=128$ here), but not so many that the model begins encoding noise ($K=512$ hurts).
  • Because the boundary-focused uncertainty loss concentrates the correction on edge pixels, the method's advantage should show up most clearly in boundary-sensitive evaluation rather than only in whole-region Dice.

Reading between the lines

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

  • The paper leaves implicit that the CIBM module is a generic per-channel weighting of decoder features, so the same module could be attached to any encoder–decoder network; a fair test of the causal claim would be to replace the learned Gaussian confounders with a fixed, random conditioning vector and check whether accuracy holds.
  • Because the posterior branch is trained on the ground-truth mask, the confounder distribution can absorb target information; a stress test would perturb or corrupt the mask at inference and check whether the learned components track the perturbation, which would reveal a target leak rather than a confounder model.
  • The framework suggests a recipe for other dense prediction tasks with known dataset biases — lesion detection, panoptic segmentation, cross-modal registration — where latent confounder modeling plus backdoor-style conditioning could replace manual bias annotation.
  • A non-Gaussian prior, such as a normalizing flow, could replace the multi-Gaussian model; if deconfounding performance survives the swap, the causal structure rather than the Gaussian form is the operative ingredient.
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

4 major / 5 minor

Summary. The paper proposes MAMBO-NET, a UNeXt-based segmentation architecture augmented by two causal-inspired components: a Gaussian self-modeling (GSm) module that learns K latent 'confusion factors' and a causal intervention for bias mitigation (CIBM) module that fuses these factors with decoder features. The central claim is that CIBM implements a backdoor adjustment, so that MAMBO-NET predicts P(Y'|do(X)) rather than the biased P(Y'|X), thereby removing the influence of confounding factors on segmentation. The authors report experiments on five datasets spanning ultrasound, dermoscopy, and colonoscopy, with Dice/IoU/FDR/AUC gains relative to several baselines, plus ablations over K, module design, and extractor backbones.

Significance. If the causal claim were valid, a general latent-modeling backdoor intervention for segmentation would be a useful and relatively novel contribution to the causal-inference branch of medical image analysis. The paper is empirically extensive: five datasets, thirteen baselines, ablations of the Gaussian component count and of backbone choices, and visualization of decoder entropy maps. However, the central derivation is invalid as written: Eq. (9) does not implement the backdoor adjustment stated in Eq. (1), and the experiments do not include uncertainty quantification or significance tests. The paper also does not provide code or reproducible evaluation scripts, and several reported numbers are inconsistent with the tables. These issues prevent the causal contribution from being assessed as stated.

major comments (4)
  1. [Sec. 3.4, Eq. (9)] The first equality in Eq. (9) is not the backdoor adjustment of Eq. (1). Eq. (1) averages over the prior P(c), whereas Eq. (9) averages over P(C|X). By the law of total expectation, E_{C~P(C|X)}[P(Y'|X,C)] = P(Y'|X), so the expression collapses to exactly the biased conditional distribution the authors seek to correct. The subsequent approximations replace the sum over confounder values with the single point feature Omega x Z (Eq. 10) and provide no argument that this preserves the causal semantics of the intervention. As written, the derivation does not establish that MAMBO-NET computes P(Y'|do(X)); the module is, at best, a feature-conditioning layer. This is the load-bearing step for the abstract and introduction claims and must be corrected or the causal claim must be substantially reframed.
  2. [Sec. 4.1, Tables 1-5] No standard deviations, number of training runs, or statistical significance tests are reported anywhere in the experiments. Several headline improvements are below one Dice point (e.g., 0.64% on ISIC2018 and 0.06% on TUI relative to UNeXt), and the K=512 ablation differs from K=128 by 0.04 Dice points. Without repeated-seed results and paired significance tests, these differences cannot be distinguished from training noise, and the statement that the method 'significantly reduces' the influence of confusion factors is not supported by the evidence presented.
  3. [Sec. 4.3, Table 3] The text states that increasing K from 128 to 512 makes Dice decline by 0.04% and IoU by 0.11%, but the table shows Dice 77.94 to 77.98 and IoU 64.49 to 64.60, i.e., slight increases. The reported saturation/decline narrative for the K ablation is contradicted by the table and must be corrected.
  4. [Sec. 4.4, Table 4] The ablation controls replace CIBM with feature concatenation and linear mapping and GSm with decoder features, which is a useful start, but they do not include a non-causal module of comparable capacity that uses the same Gaussian machinery without the do-operator framing. Because Eq. (9) does not implement an intervention, the observed Dice/FDR gains could plausibly arise from added parameters and learned feature reweighting. A direct test of confounder removal, such as a synthetic confounder experiment or a measure of independence between the learned factors and the segmentation prediction, is needed to support the causal interpretation.
minor comments (5)
  1. [Abstract and Table 2] The abstract reports FDR reductions of 1.49% (ISIC2018) and 1.87% (KVASIR), but Table 2 shows ISIC2018 FDR falling from 19.73 to 11.39 (a reduction of 8.34 percentage points) and KVASIR from 12.11 to 10.24 (1.87 percentage points). The 1.49% figure does not appear in the tables; please state the exact comparison basis for every headline number.
  2. [Sec. 3.3, Eq. (6)] The sentence 'within within the range of the boundary' contains a typo and the definition of the boundary region is incomplete; please clarify how b_i is derived from the Sobel edge map and how \hat b_i is obtained from the predicted probability map.
  3. [Sec. 3.4, Eq. (9)] The notation P(\hat C|X), \hat c_i, and the operator \odot is not formally defined; the expression x \odot \sum \hat c_i (...) mixes an image with a scalar/vector sum. A precise definition of all symbols and operations is necessary for reproducibility.
  4. [Sec. 3.3 and Sec. 4.3] The phrase 'large constant KK' should be 'large constant K', and the heading 'Analysis of K sets Gaussian Distributions' should be 'Analysis of K Gaussian Distributions'.
  5. [Sec. 4.2] The sentence 'outperforming UNeXt by 3.66% in Dice and 2.33% in IoU' is not consistent with Table 1: BUSI IoU improves by 5.21 points and DDTI IoU by 4.44 points, while no entry in Table 1 shows a 2.33-point IoU improvement. Please reconcile the reported aggregates with the table entries.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity; Eq. (9) is mathematically unsupported but not tautological, and the reported gains are held-out metric improvements.

full rationale

Walking the causal chain: Eq. (1) defines P(Y'|do(X)) by classical backdoor adjustment, and Eq. (9) re-expresses it as E_{C~P(C|X)}[P(Y'|X,C)], then collapses the sum to the point feature Ω×Z. The paper supplies no proof that this posterior substitution equals the do-intervention; in general it does not unless P(C|X)=P(C), which would contradict the paper's own confounding premise. That is a serious rigor/correctness gap, and the abstract promises to "analyze their feasibility" without providing the analysis. Under the circularity rules, however, this is not circular: no equation in the chain is identical to the quantity it is supposed to predict, nor is any reported result a refitting of its own input. Dice/IoU/FDR are held-out test metrics from five external datasets, and the model is optimized only through the stated training losses. The one self-citation in the reference list ([26], which includes author Ruiguo Yu) supplies the TUI dataset; it is a data citation and is not load-bearing for the causal mechanism. No fitted parameter is renamed as a prediction, and no uniqueness or ansatz is imported via a self-citation chain. The appropriate finding is therefore no significant circularity, with the causal-validity concern recorded as correctness risk rather than circularity; the score of 2 reflects only the minor, non-load-bearing self-citation.

Assumptions & free parameters 2 free parameters · 5 assumptions · 1 invented entities

The method rests on a set of modeling assumptions about confounders that are plausible but untested. The main contribution is empirical, but the causal interpretation depends on the validity of Eq. (9), which is not established. No machine-checked proofs or external benchmarks are provided.

free parameters (2)
  • K = 128
    Number of Gaussian components for confounder modeling, chosen by ablation on BUSI (Table 3) and used across all datasets. Performance changes with K.
  • Omega = learned (dimension n x K)
    Learnable mixing coefficients in CIBM (Eq. 10), trained end-to-end. It is a model parameter but is explicitly the mechanism for weighted fusion of confusion features.
assumptions (5)
  • domain assumption The causal graph C -> X, C -> Y', X -> Y' is correct, with C a confounder.
    Introduced in Sec. 3.2 and Fig. 3; no empirical test that this graph matches the data-generating process.
  • ad hoc to paper Confounders can be represented by K independent Gaussian latent distributions conditioned on X.
    Eq. (2) assumes Gaussian sufficiency for arbitrary abstract and concrete confounders, without justification.
  • domain assumption The posterior distribution learned from ground-truth masks (Eq. 4) provides a valid training target and can be discarded at inference.
    Relies on the variational autoencoder assumption that prior and posterior align; no test-time correction mechanism is provided.
  • ad hoc to paper The approximation in Eq. (9) is a faithful implementation of backdoor adjustment.
    Eq. (9) replaces P(C) with P(C|X) and a point-sum, which is not the standard backdoor formula; no proof of equivalence is given.
  • domain assumption Standard do-calculus assumptions hold, including no unobserved confounding beyond C.
    Needed for Eq. (1) to be valid; the paper does not discuss other potential confounders.
invented entities (1)
  • Latent confusion factors C
    purpose: Latent variables intended to represent abstract and concrete confounders such as acquisition habits, anatomy, and imaging artifacts.
    The model introduces K Gaussian latent variables with no falsifiable handle outside the paper; their existence is only inferred from segmentation metric changes.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MAMBO-NET: Multi-Causal Aware Modeling Backdoor-Intervention Optimization for Medical Image Segmentation Network." pith.science (2026). https://pith.science/paper/3NYX3KOR

@misc{pith2026250521874,
  author       = {Pith},
  title        = {Pith review of: MAMBO-NET: Multi-Causal Aware Modeling Backdoor-Intervention Optimization for Medical Image Segmentation Network},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3NYX3KOR}},
  note         = {Machine review of arXiv:2505.21874}
}
read the original abstract

Medical image segmentation methods generally assume that the process from medical image to segmentation is unbiased, and use neural networks to establish conditional probability models to complete the segmentation task. This assumption does not consider confusion factors, which can affect medical images, such as complex anatomical variations and imaging modality limitations. Confusion factors obfuscate the relevance and causality of medical image segmentation, leading to unsatisfactory segmentation results. To address this issue, we propose a multi-causal aware modeling backdoor-intervention optimization (MAMBO-NET) network for medical image segmentation. Drawing insights from causal inference, MAMBO-NET utilizes self-modeling with multi-Gaussian distributions to fit the confusion factors and introduce causal intervention into the segmentation process. Moreover, we design appropriate posterior probability constraints to effectively train the distributions of confusion factors. For the distributions to effectively guide the segmentation and mitigate and eliminate the Impact of confusion factors on the segmentation, we introduce classical backdoor intervention techniques and analyze their feasibility in the segmentation task. To evaluate the effectiveness of our approach, we conducted extensive experiments on five medical image datasets. The results demonstrate that our method significantly reduces the influence of confusion factors, leading to enhanced segmentation accuracy.

Figures

Figures reproduced from arXiv: 2505.21874 by the authors.

Figure 1
Figure 1. The segmentation results of multiple networks. The third row corre [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The architecture of the proposed MAMBO-NET. Dashed lines in￾dicate that the data stream is disabled during model inference. Segmentation Encoder and Segmentation Decoder are the encoder and the decoder in UN￾eXt. Gaussian Backbone and Posterior Constrain Backbone will use the global average pooling(GAP) and linear mapping for scale alignment. backdoor adjustment to suppress confusion effects through: ex￾plicit causa… view at source ↗
Figure 3
Figure 3. The process of causal relationship modeling. X denotes the original [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Visualization results of segmentation of multiple models on ultra [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Feature entropy map generated by the decoder layer, where [PITH_FULL_IMAGE:figures/full_fig_p006_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. CAPRI-CT: Causal Analysis and Predictive Reasoning for Image Quality Optimization in Computed Tomography

    cs.CV 2025-07 reject novelty 4.0 of 10

    CAPRI-CT predicts CT signal-to-noise ratio from images and scan metadata with a VAE ensemble, but its causal intervention and counterfactual claims rest on an unjustified identifiability assumption.

Reference graph

Works this paper leans on

40 extracted references · 36 canonical work pages · cited by 1 Pith paper

  1. [1]

    Ronneberger, P

    O. Ronneberger, P. Fischer, T. Brox, U-net: Convolutional networks for biomedical image segmentation, in: Medi- cal image computing and computer-assisted intervention– MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, Springer, 2015, pp. 234–241

    2015

  2. [2]

    Zhang, Q

    Z. Zhang, Q. Liu, Y . Wang, Road extraction by deep resid- ual u-net, IEEE Geoscience and Remote Sensing Letters 15 (5) (2018) 749–753

    work page 2018

  3. [3]

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

    work page 2018

  4. [4]

    J. M. J. Valanarasu, V . M. Patel, Unext: Mlp-based rapid medical image segmentation network, in: International conference on MICCAI, Springer, 2022, pp. 23–33

    work page 2022

  5. [5]

    Hernan, J

    M. Hernan, J. Robins, Causal inference: What if chapman hall/crc, boca raton (2020)

    work page 2020

  6. [6]

    N. Zhao, A. Basarab, D. Kouamé, J.-Y . Tourneret, Joint segmentation and deconvolution of ultrasound images us- ing a hierarchical bayesian model based on generalized gaussian priors, IEEE transactions on Image Processing 25 (8) (2016) 3736–3750

    work page 2016

  7. [7]

    Z. Zhao, S. Du, Z. Xu, Z. Yin, X. Huang, X. Huang, C. Wong, Y . Liang, J. Shen, J. Wu, J. Qu, L. Zhang, Y . Cui, Y . Wang, L. Wee, A. Dekker, C. Han, Z. Liu, Z. Shi, C. Liang, Swinhr: Hemodynamic-powered hierar- chical vision transformer for breast tumor segmentation, Computers in Biology and Medicine 169 (2024) 107939

    work page 2024

  8. [8]

    B. Liu, D. Wang, X. Yang, Y . Zhou, R. Yao, Z. Shao, J. Zhao, Show, deconfound and tell: Image captioning with causal inference, in: Proceedings of the IEEE/CVF Conference on CVPR, 2022, pp. 18041–18050

    work page 2022

Show all 40 references
  1. [9]

    X. Yang, H. Zhang, J. Cai, Deconfounded image caption- ing: A causal retrospect, IEEE Transactions on Pattern Analysis and Machine Intelligence 45 (11) (2021) 12996– 13010

    2021

  2. [10]

    J. Wang, C. Zhong, C. Feng, Y . Zhang, J. Sun, Y . Yokota, Disentangled representation for cross-domain medical im- age segmentation, IEEE Transactions on Instrumentation and Measurement 72 (2022) 1–15

    2022

  3. [11]

    R. Gu, G. Wang, J. Lu, J. Zhang, W. Lei, Y . Chen, W. Liao, S. Zhang, K. Li, D. N. Metaxas, et al., Cddsa: Contrastive domain disentanglement and style augmentation for gen- eralizable medical image segmentation, Medical Image Analysis 89 (2023) 102904

    2023

  4. [12]

    X. Li, Y . Pang, Y . Yuan, L1-norm-based 2dpca, IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics) 40 (4) (2010) 1170–1175

    2010

  5. [13]

    Y . Wang, B. Xiao, X. Bi, W. Li, X. Gao, Boundary-aware prototype in semi-supervised medical image segmenta- tion, IEEE Transactions on Image Processing (2024)

    2024

  6. [14]

    Y . Liu, L. Lin, K. K. Wong, X. Tang, Procns: Progressive prototype calibration and noise suppression for weakly- supervised medical image segmentation, IEEE Journal of Biomedical and Health Informatics (2024)

    2024

  7. [15]

    L. Zhou, L. Wang, W. Li, B. Lei, J. Mi, W. Yang, Multi- stage liver segmentation in ct scans using gaussian pseudo variance level set, IEEE Access 9 (2021) 101414–101423

    2021

  8. [16]

    S. Gao, H. Zhou, Y . Gao, X. Zhuang, Bayeseg: Bayesian modeling for medical image segmentation with inter- pretable generalizability, Medical Image Analysis 89 (2023) 102889

    2023

  9. [17]

    Z. Chen, Z. Tian, J. Zhu, C. Li, S. Du, C-cam: Causal cam for weakly supervised semantic segmentation on medical image, in: Proceedings of the IEEE/CVF Conference on CVPR, 2022, pp. 11676–11685

    2022

  10. [18]

    X. Song, J. Liu, Y . Liu, Y . Li, W. Lei, R. Wang, Re- thinking radiology report generation via causal inspired counterfactual augmentation, in: Proceedings of the 15th ACM International Conference on Bioinformatics, Com- putational Biology and Health Informatics, 2024, pp. 1– 10

    2024

  11. [19]

    K. Chen, S. Sun, J. Zhao, Camil: Causal multiple instance learning for whole slide image classification, in: Proceed- ings of the AAAI Conference on Artificial Intelligence, V ol. 38, 2024, pp. 1120–1128

    2024

  12. [20]

    Y . Chen, M. Wei, Z. Zheng, J. Hu, Y . Shi, S. Xiong, X. X. Zhu, L. Mou, Causalclipseg: Unlocking clip’s potential in referring medical image segmentation with causal intervention, in: International Conference on Med- ical Image Computing and Computer-Assisted Interven- tion, ...

    2024

  13. [21]

    C. Wang, J. Li, F. Zhang, X. Sun, H. Dong, Y . Yu, Bilateral asymmetry guided counterfactual generating network for mammogram classification, IEEE Transactions on Image Processing 30 (2021) 7980–7994

    2021

  14. [22]

    J. G. Richens, Improving the accuracy of medical diagno- sis with causal machine learning, Nature communications 11 (1) (2020) 3923

    2020

  15. [23]

    D. P. Kingma, Auto-encoding variational bayes, arXiv preprint arXiv:1312.6114 (2013)

    2013 arXiv

  16. [24]

    Al-Dhabyani, M

    W. Al-Dhabyani, M. Gomaa, H. Khaled, A. Fahmy, Dataset of breast ultrasound images, Data in brief 28 (2020) 104863. 7

    2020

  17. [25]

    Pedraza, C

    L. Pedraza, C. Vargas, F. Narváez, O. Durán, E. Muñoz, E. Romero, An open access thyroid ultrasound image database, in: 10th International symposium on medical in- formation processing and analysis, V ol. 9287, SPIE, 2015, pp. 188–193

    2015

  18. [26]

    Z. Bai, L. Chang, R. Yu, X. Li, X. Wei, M. Yu, Z. Liu, J. Gao, J. Zhu, Thyroid nodules risk stratification through deep learning based on ultrasound images, Medi- cal Physics 47 (12) (2020) 6355–6365

    2020

  19. [27]

    M. A. Al-Masni, D.-H. Kim, Multiple skin lesions di- agnostics via integrated deep convolutional networks for segmentation and classification, Computer methods and programs in biomedicine 190 (2020) 105351

    2020

  20. [28]

    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: MultiMedia model- ing: 26th international conference, MMM 2020, Daejeon, South Korea, January 5–8, 2020, proceedings, part II 26, Springer...

    2020

  21. [29]

    J. M. J. Valanarasu, P. Oza, I. Hacihaliloglu, V . M. Pa- tel, Medical transformer: Gated axial-attention for medi- cal image segmentation, in: MICCAI 2021: 24th interna- tional conference, Springer, 2021, pp. 36–46

    2021

  22. [30]

    C. Xue, L. Zhu, H. Fu, X. Hu, X. Li, H. Zhang, P. A. Heng, Global guidance network for breast lesion segmentation in ultrasound images, Medical Image Analysis 70 (2021) 101989

    2021

  23. [31]

    J. Chen, Y . Lu, Q. Yu, X. Luo, E. Adeli, Y . Wang, L. Lu, A. L. Yuille, Y . Zhou, Transunet: Transformers make strong encoders for medical image segmentation, arXiv preprint arXiv:2102.04306 (2021)

    2021 arXiv

  24. [32]

    G. Chen, L. Li, Y . Dai, J. Zhang, M. H. Yap, Aau-net: an adaptive attention u-net for breast lesions segmentation in ultrasound images, IEEE Transactions on Medical Imag- ing 42 (5) (2022) 1289–1300

    2022

  25. [33]

    H. Cao, Y . Wang, J. Chen, D. Jiang, X. Zhang, Q. Tian, Swin-unet: Unet-like pure transformer for medical image segmentation, in: European conference on computer vi- sion, Springer, 2022, pp. 205–218

    2022

  26. [34]

    Ibtehaz, D

    N. Ibtehaz, D. Kihara, Acc-unet: A completely convolu- tional unet model for the 2020s, in: International Con- ference on Medical Image Computing and Computer- Assisted Intervention, Springer, 2023, pp. 692–702

    2023

  27. [35]

    H. Wu, X. Huang, X. Guo, Z. Wen, J. Qin, Cross-image dependency modeling for breast ultrasound segmentation, IEEE Transactions on Medical Imaging 42 (06) (2023) 1619–1631

    2023

  28. [36]

    Vezakis, K

    I. Vezakis, K. Georgas, D. Fotiadis, G. Matsopoulos, Eff- isegnet: Gastrointestinal polyp segmentation through a pre-trained efficientnet-based network with a simplified decoder, in: Proceedings of the International Conference of the IEEE Engineering in Medicine and Biology So...

    2024

  29. [37]

    Z. Lu, C. She, W. Wang, Q. Huang, Lm-net: A light- weight and multi-scale network for medical image seg- mentation, Computers in Biology and Medicine 168 (2024) 1–12

    2024

  30. [38]

    Simonyan, Very deep convolutional networks for large- scale image recognition, arXiv preprint arXiv:1409.1556 (2014)

    K. Simonyan, Very deep convolutional networks for large- scale image recognition, arXiv preprint arXiv:1409.1556 (2014)

    2014 arXiv

  31. [39]

    Huang, Z

    G. Huang, Z. Liu, L. Van Der Maaten, K. Q. Weinberger, Densely connected convolutional networks, in: Proceed- ings of the IEEE conference on CVPR, 2017, pp. 4700– 4708

    2017

  32. [40]

    K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in: Proceedings of the IEEE con- ference on CVPR, 2016, pp. 770–778. 8

    2016

Pith tools

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