Subtyping Breast Lesions via Generative Augmentation based Long-tailed Recognition in Ultrasound

Ao Chang; Dong Ni; Ruobing Huang; Shijing Chen; Xinrui Zhou; Yuhao Huang; Yuhao Wang

REVIEW 4 major objections 5 minor 31 references

Subtyping Breast Lesions via Generative Augmentation based Long-tailed Recognition in Ultrasound

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Breast-lesion subtype classification in ultrasound improves when a sketch-guided diffusion synthesizer is paired with reinforcement-learning agents that decide how many synthetic images go into each training batch.

desk verdict A useful training-recipe paper with a promising direction; the missing statistics and garbled REINFORCE equation keep the claims from being solid yet. read the letter →

arxiv 2507.22568 v1 pith:3OMD3VLZ submitted 2025-07-30 cs.CV

Shijing Chen , Xinrui Zhou , Yuhao Wang , Yuhao Huang , Ao Chang , Dong Ni , Ruobing Huang This is my paper

classification cs.CV

keywords breastultrasoundlong-tailedrecognitiongenerativeaugmentationlatentdiffusionmodelsreinforcementlearningsketchsupervisionhistologicalsubtypeclassification

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

Breast lesion subtypes appear in ultrasound with a severely long-tailed distribution, and the tail classes are exactly the ones that matter for personalized treatment. The paper argues that generative augmentation can fix this imbalance, but only when the synthetic images are both anatomically faithful and used sparingly: it proposes a two-phase framework that first synthesizes class-conditioned ultrasound images with a latent diffusion model guided by a sketch-based structural loss, then trains the classifier on batches whose synthetic-real ratio is chosen per class by reinforcement-learning agents. On an in-house 8-class dataset (Breast-LT-8, imbalance ratio 47.98:1) the method reports F1=35.23, all accuracy=72.31, and few-shot accuracy=31.25, outperforming the best generative baseline Skin-SDM (F1=33.30, all=68.59, few=9.38); on public BreastMNIST it reports accuracy 89.10. If this holds, it suggests that success in long-tailed medical classification depends as much on how synthetic data are deployed as on how they are generated.

What carries the argument

The machinery is the class-steerable synthesizer plus the RL-driven class adaptive sampler (RL-CAS). The synthesizer is a latent diffusion model conditioned on the subtype label $c$; its denoiser has a sketch-grounded perception branch that takes multi-scale latent features from the encoder path, decodes them into a predicted sketch $S_{pred}$, and is trained with the auxiliary loss $\mathcal{L}_s = \sqrt{\bar\alpha_t}\, L_1(S_{pred}, S_{gt})$, where $S_{gt}$ is a sketch precomputed by a pixel-difference edge detector. The RL-CAS treats each class as an agent that changes that class's synthetic-image count in a mini-batch by an action in $\{-2,0,+2\}$; after $K$ parallel classifier trajectories, the agents are updated with the REINFORCE rule using reward $(\epsilon_j + 0.04)^3$ and a moving baseline $B_t$. These two pieces do the paper's work: the sketch loss keeps synthetic images anatomically faithful, and the sampler keeps synthetic data from overwhelming the classifier while still addressing tail scarcity.

What would settle it

Run the full pipeline on Breast-LT-8 with the sketch loss removed but RL-CAS kept; if few-shot accuracy does not fall below 31.25, the paper's attribution of tail-class gains to sketch-grounded structural supervision is falsified.

Watch

Extended reading notes

Core claim

The central claim is that the long-tailed class imbalance in breast ultrasound subtype classification can be materially reduced by correcting the data distribution in two coordinated ways: synthesize class-conditioned ultrasound images with a latent diffusion model whose denoiser is additionally trained to reproduce a sketch of the lesion, and then let per-class reinforcement-learning agents set how many synthetic images each training mini-batch contains. The paper reports this dual-phase framework reaches F1=35.23, all-shot accuracy=72.31, and few-shot accuracy=31.25 on the in-house Breast-LT-8 dataset, and accuracy 89.10 on BreastMNIST. It also claims the sketch branch keeps the synthetic images diagnostically faithful while requiring no test-time annotations, since the sketch is only computed during training.

Load-bearing premise

The synthesis stage trusts that a VAE pretrained on natural images and left unfine-tuned on ultrasound can faithfully capture speckle and lesion texture in its latent space; if it cannot, the generated training images will not carry the class-discriminative structure the classifier is supposed to learn from.

Editorial extensions

If this is right

Generative augmentation alone gives only a small rise in F1 (30.94 to 31.82); the sketch branch adds the next increment (to 34.73), indicating the paper sees structural fidelity as a principal driver of tail-class gains.
Static re-sampling of synthetic data lowers F1 to 29.48, so the paper claims the RL-CAS is what prevents synthetic overuse from hurting overall accuracy.
On BreastMNIST the framework reports 89.10 accuracy, extending the approach to a public binary-imbalance setting, though the gaps to baselines shrink there.
At test time the classifier is unchanged from a standard backbone; all data synthesis and sampling decisions happen during training, preserving fast inference.

Reading between the lines

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

The reported improvements are small in absolute terms, so a natural extension is to rerun the framework over multiple random splits and report variance; without that, the margin over Skin-SDM (F1 35.23 vs 33.30) cannot be told apart from run-to-run noise.
Because the synthesis stage leaves the pretrained VAE frozen, an obvious test is fine-tuning the VAE on ultrasound before training the denoiser; if tail-class F1 does not improve further, the pretrained latent space was not the bottleneck.
The same RL-CAS could be applied to other long-tailed medical imaging tasks, but the reward function and action step size would likely need re-tuning to the target validation metric.

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.

Desk Editor's Note

read the letter

Short version: this is a plausible recipe paper for long-tailed breast-ultrasound classification, but the main evidence isn't built to carry the claim yet. The combination—class-conditioned latent diffusion with a sketch structural loss, plus a multi-agent REINFORCE sampler that adjusts synthetic/real batch composition—is new as a package, and the ablations are consistent. The clinical task is real and the comparison table is serious. That's the good part.

The soft spots are concentrated in the evaluation and the RL derivation. Every number in Table 1 comes from a single run on a private dataset, with no per-class test counts and no error bars. The Few-shot gain over Skin-SDM (31.25 vs 9.38) could be a difference of a few images, and the RL-CAS selecting the best of K=3 validation classifiers per epoch adds selection bias on top. This is load-bearing; the stress-test note is right. Also, Eq. (1) is wrong as written: it gives agent i an update computed from the gradient of the joint log-probability Σ_i log p(a_i) with respect to θ, not the agent's own θ_i. Could be a typo, but it needs fixing. The reward metric is never named. Minor issue: the text says 'across all metrics,' but Table 1 shows AUC below Skin-SDM and MGDM. The VAE-transferability concern is minor; a suboptimal pre-trained compressor would hurt fidelity, but that is secondary to the missing statistics. No code or dataset release beyond sample images.

Who this is for: people building long-tailed medical classifiers or using generative augmentation. It deserves a serious referee because the recipe is interesting and the task is real, but I would not cite the numbers as they stand. My recommendation for review: send it, but expect major revision—add seeds and error bars, report per-class counts, correct the RL update, and soften the superiority claim.

Referee Report

4 major / 5 minor

Summary. The paper proposes a dual-phase framework for long-tailed breast lesion subtype classification in ultrasound. The first phase is a class-conditioned latent diffusion synthesizer augmented with a sketch-grounded perception branch that adds an L1 sketch loss to the standard LDM objective. The second phase is an RL-driven class adaptive sampler (RL-CAS) that, over K episodes per training epoch, adjusts the number of synthetic samples per class in mini-batches and uses a REINFORCE-style update with a cubic reward based on a validation metric. The authors evaluate on a private in-house Breast-LT-8 dataset (8 subtypes, 5622 images, imbalance ratio 47.98:1) and on public BreastMNIST, reporting F1=35.23, All accuracy=72.31, Few accuracy=31.25 on Breast-LT-8 and accuracy=89.10 on BreastMNIST, outperforming re-balancing, architecture-enhanced, and generative augmentation baselines. The central claim is that the combination of high-fidelity synthetic data and adaptive synthetic-real ratio calibration yields promising long-tailed recognition performance.

Significance. If the empirical claims are reliable, the framework addresses a clinically relevant problem: breast lesion subtype classification under heavy class imbalance. The idea of coupling a class-conditioned synthesizer with an RL-driven sampler is reasonable and goes beyond simple generative oversampling, which is often detrimental when applied indiscriminately. The paper also provides a potentially reusable design by separating synthesis from sampling and by using sketch-based structural supervision that does not require test-time annotations. However, the significance is currently tempered by the absence of uncertainty quantification: all headline numbers come from single runs on a private dataset, and the main SOTA gaps are small in F1 terms. The paper does not provide code, per-class test counts, confidence intervals, or repeated-seed experiments, so the reader cannot currently distinguish a genuine improvement from sampling noise or selection bias. The work is also not yet accompanied by a machine-checked derivation or a parameter-free construction; the RL update in Eq.

major comments (4)

[§2.2, Eq. (1)]
[§3, Table 1]
[§2.1, §3]
[§2.2, Eq. (2)]

minor comments (5)

[§3, Method Comparison]
[§3, Table 1]
[§2.2]
[§2.1]
[§3]

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's claimed gains are empirical benchmark results, and no reported metric is equal by construction to a fitted input.

full rationale

Walking the derivation chain, the framework's components are standard generative and RL objectives: the LDM loss in Section 2.1, the sketch loss Ls = sqrt(alpha_bar_t) * L1(Spred, Sgt), and the REINFORCE update in Eq. (1). None of these equations define the reported F1/accuracy values in terms of themselves; the target results come from held-out test evaluations (Breast-LT-8 test split and BreastMNIST). The RL-CAS does use a validation metric as its reward and selects the best of K=3 classifiers per epoch on validation, but this is checkpoint/model selection on a held-out split, not fitting to the test labels used for Table 1. Self-citations to [30] (class-label injection design), [22] (baseline term), and [31] (clinical annotation availability) support implementation details only; they do not carry the central claim and no uniqueness theorem is imported. The transferability of the pretrained VAE and sketch extractor to ultrasound is an empirical assumption, not a circularity. The public BreastMNIST benchmark and comparisons with external SOTA baselines give the claim independent content. Therefore no step reduces to its own input.

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

The central claim does not rest on a derivation, but on an experimental pipeline. The load-bearing premises are the transferability of a pre-trained VAE and sketch extractor to ultrasound, and the assumptions encoded in the RL sampler (reward design, validation-based selection). These are summarized below.

free parameters (5)

sketch loss weight lambda = 0.1
Controls the strength of the structural perception branch; no sensitivity analysis is reported.
RL episodes per epoch K = 3
Number of candidate batch compositions and parallel classifiers trained each epoch; chosen by hand.
state initialization s_i = 2
Initial synthetic sample count per class in a mini-batch; set to 2 without stated reasoning.
action step size = [-2, 0, +2]
Discrete action set for each agent adjusting per-class synthetic counts; no grid search reported.
reward hyperparameters = R=(epsilon+0.04)^3, gamma=0.99
Cubic compression and shift offset are ad hoc choices to amplify reward signal; no ablation on these constants.

assumptions (5)

standard math DDPM/LDM training objective and REINFORCE gradient estimator are valid
The synthesizer uses the LDM objective (Eq. L_LDM) and the sampler uses the REINFORCE rule (Eq. 1); the paper assumes these are correct and implementable as stated.
domain assumption Pre-trained VAE and sketch extractor transfer to breast ultrasound
z0 is encoded with a pre-trained VAE [7] and S_gt is computed with a pre-trained edge detector [20]; no fine-tuning of these modules on ultrasound is reported in Section 2.1 or Section 3.
domain assumption Class label conditioning is sufficient for subtype-specific synthesis
The synthesizer conditions only on the subtype label (c) and the sketch loss; the paper assumes this captures the morphological differences between subtypes.
domain assumption Validation metric used to select the best classifier generalizes to the test set
RL-CAS selects the classifier with the highest validation metric each epoch; the paper assumes this selector does not overfit the validation set in a way that inflates test results.
ad hoc to paper Cubic reward function improves training
The reward R=(epsilon+0.04)^3 is introduced to 'enhance the reward signal' without a derivation or sensitivity study; it is specific to this paper's training setup.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Subtyping Breast Lesions via Generative Augmentation based Long-tailed Recognition in Ultrasound." pith.science (2026). https://pith.science/paper/3OMD3VLZ

@misc{pith2026250722568,
  author       = {Pith},
  title        = {Pith review of: Subtyping Breast Lesions via Generative Augmentation based Long-tailed Recognition in Ultrasound},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3OMD3VLZ}},
  note         = {Machine review of arXiv:2507.22568}
}

read the original abstract

Accurate identification of breast lesion subtypes can facilitate personalized treatment and interventions. Ultrasound (US), as a safe and accessible imaging modality, is extensively employed in breast abnormality screening and diagnosis. However, the incidence of different subtypes exhibits a skewed long-tailed distribution, posing significant challenges for automated recognition. Generative augmentation provides a promising solution to rectify data distribution. Inspired by this, we propose a dual-phase framework for long-tailed classification that mitigates distributional bias through high-fidelity data synthesis while avoiding overuse that corrupts holistic performance. The framework incorporates a reinforcement learning-driven adaptive sampler, dynamically calibrating synthetic-real data ratios by training a strategic multi-agent to compensate for scarcities of real data while ensuring stable discriminative capability. Furthermore, our class-controllable synthetic network integrates a sketch-grounded perception branch that harnesses anatomical priors to maintain distinctive class features while enabling annotation-free inference. Extensive experiments on an in-house long-tailed and a public imbalanced breast US datasets demonstrate that our method achieves promising performance compared to state-of-the-art approaches. More synthetic images can be found at https://github.com/Stinalalala/Breast-LT-GenAug.

Figures

Figures reproduced from arXiv: 2507.22568 by the authors.

**Figure 1.** Breast US images of lesions with different histological subtypes. The histogram indicates the incidence of different subtypes, which exhibit a long-tailed distribution. Red-bordered images are real US images, while the blue-bordered ones are synthetic data generated using the proposed framework. emphasize binary lesion classification regarding malignancy/benignity, contemporary clinical findings suggested that the … view at source ↗

**Figure 2.** Pipeline of our proposed framework. E, D indicate pre-trained encoder and decoder in VAE, respectively. zt refers to latent features after the t-step diffusion process. RL, reinforcement learning. through several convolution-attention hybrid blocks. To ensure diagnostic relevance, we implement disease-specific generation control by injecting class labels into the denoising trajectory, following a similar approach … view at source ↗

**Figure 3.** Illustration of the RL-driven class adaptive sampler. Starting from a uniform initial state (green histogram on the left), multiple agents take different actions to modify the ratio of synthetic data during training to optimize the reward. The core idea is to optimize the high-dimensional feature space of the denoiser in a self-supervised manner through an L1 loss to minimize the reconstruction error between the pre… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 21 canonical work pages

[1]

The Breast 62, S3–S6 (2022)

Barrios, C.H.: Global challenges in breast cancer detection and treatment. The Breast 62, S3–S6 (2022)

work page 2022
[2]

The British journal of radiology91(1083), 20170576 (2018)

Becker, A.S., Mueller, M., Stoffel, E., Marcon, M., Ghafoor, S., Boss, A.: Classifica- tion of breast cancer in ultrasound imaging using a generic deep learning analysis software: a pilot study. The British journal of radiology91(1083), 20170576 (2018)

work page 2018
[3]

Cui, Y., Jia, M., Lin, T.Y., Song, Y., Belongie, S.: Class-balanced loss based on effective number of samples (2019)

work page 2019
[4]

IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)

Du, C., Wang, Y., Song, S., Huang, G.: Probabilistic contrastive learning for long- tailed visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)

work page 2024
[5]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

Du, F., Yang, P., Jia, Q., Nan, F., Chen, X., Yang, Y.: Global and local mixture consistency cumulative learning for long-tailed visual recognitions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15814–15823 (2023)

work page 2023
[6]

In: Robertson, D., Williams, G.H

Dunne, R.M., O’Neill, A.C., Tempany, C.M.: Chapter 9 - imaging tools in clin- ical research: Focus on imaging technologies. In: Robertson, D., Williams, G.H. (eds.) Clinical and Translational Science (Second Edition), pp. 157–179. Academic Press, second edition edn. (2017).https://doi.org/https://doi.org/10.1016/ B978-0-12-802101-9.00009-0

work page 2017
[7]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12873–12883 (2021)

2021
[8]

Physics in Medicine & Biology62(19), 7714 (2017)

Han, S., Kang, H.K., Jeong, J.Y., Park, M.H., Kim, W., Bang, W.C., Seong, Y.K.: A deep learning framework for supporting the classification of breast lesions in ultrasound images. Physics in Medicine & Biology62(19), 7714 (2017)

work page 2017

Show all 31 references

[9]

Advances in neural information processing systems33, 6840–6851 (2020)

Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems33, 6840–6851 (2020)

2020
[10]

In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition

Li, J., Tan, Z., Wan, J., Lei, Z., Guo, G.: Nested collaborative learning for long- tailed visual recognition. In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition. pp. 6949–6958 (2022) 10 S. Chen and X. Zhou et al

2022
[11]

IEEE Transactions on Pattern Anal- ysis and Machine Intelligence (2024)

Luo, Y., Yang, Q., Fan, Y., Qi, H., Xia, M.: Measurement guidance in diffusion models: Insight from medical image synthesis. IEEE Transactions on Pattern Anal- ysis and Machine Intelligence (2024)

2024
[12]

Clinical medicine insights: Pathology8, CPath–S31563 (2015)

Makki, J.: Diversity of breast carcinoma: histological subtypes and clinical rele- vance. Clinical medicine insights: Pathology8, CPath–S31563 (2015)

2015
[13]

arXiv preprint arXiv:2007.07314 (2020)

Menon, A.K., Jayasumana, S., Rawat, A.S., Jain, H., Veit, A., Kumar, S.: Long-tail learning via logit adjustment. arXiv preprint arXiv:2007.07314 (2020)

2020 arXiv
[14]

Computer methods and programs in biomedicine190, 105361 (2020)

Moon,W.K.,Lee,Y.W.,Ke,H.H.,Lee,S.H.,Huang,C.S.,Chang,R.F.:Computer- aided diagnosis of breast ultrasound images using ensemble learning from convo- lutional neural networks. Computer methods and programs in biomedicine190, 105361 (2020)

2020
[15]

In: Proceedings of the AAAI conference on artificial intelligence

Mou, C., Wang, X., Xie, L., Wu, Y., Zhang, J., Qi, Z., Shan, Y.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In: Proceedings of the AAAI conference on artificial intelligence. pp. 4296– 4304 (2024)

2024
[16]

In: 2024 16th International Conference on Knowledge and Smart Technology (KST)

Patcharapimpisut, P., Khanarsa, P.: Generating synthetic images using stable dif- fusion model for skin lesion classification. In: 2024 16th International Conference on Knowledge and Smart Technology (KST). pp. 184–189 (2024)

2024
[17]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684–10695 (2022)

2022
[18]

Shi, J.X., Wei, T., Zhou, Z., Shao, J.J., Han, X.Y., Li, Y.F.: Long-tail learning with foundation model: Heavy fine-tuning hurts (2024)

2024
[19]

In: International conference on machine learning

Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsuper- vised learning using nonequilibrium thermodynamics. In: International conference on machine learning. pp. 2256–2265. PMLR (2015)

2015
[20]

In: Proceedings of the IEEE/CVF international conference on computer vision

Su, Z., Liu, W., Yu, Z., Hu, D., Liao, Q., Tian, Q., Pietikäinen, M., Liu, L.: Pixel difference networks for efficient edge detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5117–5127 (2021)

2021
[21]

CA: a cancer journal for clinicians 71(3), 209–249 (2021)

Sung, H., Ferlay, J., Siegel, R.L., Laversanne, M., Soerjomataram, I., Jemal, A., Bray, F.: Global cancer statistics 2020: Globocan estimates of incidence and mor- tality worldwide for 36 cancers in 185 countries. CA: a cancer journal for clinicians 71(3), 209–249 (2021)

2021
[22]

Medical Image Analysis77, 102362 (2022)

Wang, K.N., Yang, X., Miao, J., Li, L., Yao, J., Zhou, P., Xue, W., Zhou, G.Q., Zhuang, X., Ni, D.: Awsnet: An auto-weighted supervision attention network for myocardial scar and edema segmentation in multi-sequence cardiac magnetic res- onance images. Medical Image Analysis77...

2022
[23]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

Wang, Y., Gao, R., Chen, K., Zhou, K., Cai, Y., Hong, L., Li, Z., Jiang, L., Yeung, D.Y., Xu, Q., et al.: Detdiffusion: Synergizing generative and perceptive models for enhanced data generation and perception. In: Proceedings of the IEEE/CVF Conference on Computer Vision and P...

2024
[24]

Machine learning8, 229–256 (1992)

Williams, R.J.: Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine learning8, 229–256 (1992)

1992
[25]

Advances in Neural Information Processing Systems36, 54683–54695 (2023)

Wu, W., Zhao, Y., Chen, H., Gu, Y., Zhao, R., He, Y., Zhou, H., Shou, M.Z., Shen, C.: Datasetdm: Synthesizing data with perception annotations using diffu- sion models. Advances in Neural Information Processing Systems36, 54683–54695 (2023)

2023
[26]

Scientific Data 10(1), 41 (2023) BreastGenAug 11

Yang,J.,Shi,R.,Wei,D.,Liu,Z.,Zhao,L.,Ke,B.,Pfister,H.,Ni,B.:Medmnistv2- a large-scale lightweight benchmark for 2d and 3d biomedical image classification. Scientific Data 10(1), 41 (2023) BreastGenAug 11

2023
[27]

Yang, L., Xu, X., Kang, B., Shi, Y., Zhao, H.: Freemask: Synthetic images with dense annotations make stronger segmentation models (2023)

2023
[28]

In: Proceedings of the IEEE/CVF international conference on computer vision

Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3836–3847 (2023)

2023
[29]

IEEE Transactions on Pattern Analysis and Machine Intelligence45(9), 10795–10816 (2023)

Zhang, Y., Kang, B., Hooi, B., Yan, S., Feng, J.: Deep long-tailed learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence45(9), 10795–10816 (2023)

2023
[30]

arXiv preprint arXiv:2409.17091 (2024)

Zhou, X., Huang, Y., Dou, H., Chen, S., Chang, A., Liu, J., Long, W., Zheng, J., Xu, E., Ren, J., et al.: Ctrl-genaug: Controllable generative augmentation for medical sequence classification. arXiv preprint arXiv:2409.17091 (2024)

2024
[31]

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

Zhou, X., Huang, Y., Xue, W., Dou, H., Cheng, J., Zhou, H., Ni, D.: Heartbeat: To- wards controllable echocardiography video synthesis with multimodal conditions- guided diffusion models. In: International Conference on Medical Image Computing and Computer-Assisted Interventio...

2024

Pith tools

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

[1] [1]

The Breast 62, S3–S6 (2022)

Barrios, C.H.: Global challenges in breast cancer detection and treatment. The Breast 62, S3–S6 (2022)

work page 2022

[2] [2]

The British journal of radiology91(1083), 20170576 (2018)

Becker, A.S., Mueller, M., Stoffel, E., Marcon, M., Ghafoor, S., Boss, A.: Classifica- tion of breast cancer in ultrasound imaging using a generic deep learning analysis software: a pilot study. The British journal of radiology91(1083), 20170576 (2018)

work page 2018

[3] [3]

Cui, Y., Jia, M., Lin, T.Y., Song, Y., Belongie, S.: Class-balanced loss based on effective number of samples (2019)

work page 2019

[4] [4]

IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)

Du, C., Wang, Y., Song, S., Huang, G.: Probabilistic contrastive learning for long- tailed visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)

work page 2024

[5] [5]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

Du, F., Yang, P., Jia, Q., Nan, F., Chen, X., Yang, Y.: Global and local mixture consistency cumulative learning for long-tailed visual recognitions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15814–15823 (2023)

work page 2023

[6] [6]

In: Robertson, D., Williams, G.H

Dunne, R.M., O’Neill, A.C., Tempany, C.M.: Chapter 9 - imaging tools in clin- ical research: Focus on imaging technologies. In: Robertson, D., Williams, G.H. (eds.) Clinical and Translational Science (Second Edition), pp. 157–179. Academic Press, second edition edn. (2017).https://doi.org/https://doi.org/10.1016/ B978-0-12-802101-9.00009-0

work page 2017

[7] [7]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12873–12883 (2021)

2021

[8] [8]

Physics in Medicine & Biology62(19), 7714 (2017)

Han, S., Kang, H.K., Jeong, J.Y., Park, M.H., Kim, W., Bang, W.C., Seong, Y.K.: A deep learning framework for supporting the classification of breast lesions in ultrasound images. Physics in Medicine & Biology62(19), 7714 (2017)

work page 2017