Few-Shot Learning with Adaptive Weight Masking in Conditional GANs

Jiacheng Hu; Jiajing Chen; Jianjun Wei; Runyuan Bao; Xinyu Qiu; Zhen Qi

REVIEW 5 major objections 5 minor 1 cited by

Few-Shot Learning with Adaptive Weight Masking in Conditional GANs

T0 review · 5 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read Weight-masked GAN improves few-shot sample quality on MNIST

desk verdict Claims few-shot accuracy gains but never measures accuracy; the reported FID numbers actually work against the method, and the core equations are garbled. read the letter →

arxiv 2412.03105 v1 pith:HF7M5UAM submitted 2024-12-04 cs.CV cs.LG

Jiacheng Hu , Zhen Qi , Jianjun Wei , Jiajing Chen , Runyuan Bao , Xinyu Qiu This is my paper · ORCID

classification cs.CVcs.LG

keywords few-shotlearningconditionalGANdataaugmentationresidualnetworkweightmaskMNISTInceptionScoreFréchetDistance

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

This paper proposes RWM-CGAN, a conditional GAN for few-shot data augmentation that adds residual blocks to the generator and a weight-mask regularizer to the discriminator. The goal is to generate samples that are clearer and more diverse than a plain CGAN when training data per class is scarce, so that downstream models generalize better. On MNIST, the paper reports higher Inception Scores and lower Fréchet Inception Distances for most digit classes, with an ablation indicating both components contribute. The paper also claims improved detection and classification accuracy, though no such experiments appear in the full text.

What carries the argument

The central objects are the residual block in the generator and the weight mask in the discriminator. A residual block contains two 3x3 convolutional layers with a skip connection and batch normalization, and a 1x1 convolutional shortcut when dimensions change. The weight mask is computed by averaging the absolute difference images between qualifying training samples and a CGAN template, then normalizing; the mask is applied to discriminator weights so that low-weight interference points are suppressed. The paper's argument is that the mask lets the discriminator attend to genuine class-specific differences rather than reconstruction noise.

What would settle it

Train RWM-CGAN on a small per-class MNIST subset and then train a classifier on real plus generated samples; if test accuracy does not exceed a classifier trained with simple geometric augmentations on the same real samples, the paper's downstream claim is falsified. More directly, one can inspect the discriminator gradients with and without the mask: if the mask does not change gradient directions for the small-sample classes, the proposed mechanism is inert.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that combining residual units in the generator with a weight mask in the discriminator improves conditional GAN-based data augmentation for few-shot settings. The residual blocks deepen the generator and improve image quality; the weight mask, built by averaging difference images between training samples and a CGAN template, suppresses interference and focuses the discriminator on small-sample category features. The reported MNIST results show mean IS rising from 6.415 to 6.966 and mean FID falling from 13.721 to 10.126, with per-class FID lower for nine of ten classes. The ablation attributes gains to both modifications, though the mask-only variant achieves a better FID (9.174) than the full model (10.126).

Load-bearing premise

The load-bearing premise is that the averaged difference-image mask actually steers the discriminator toward better small-sample feature learning; the paper does not validate this mechanism experimentally, and the full model's FID is worse than the mask-only variant.

Editorial extensions

If this is right

If RWM-CGAN works as claimed, few-shot learners can obtain richer training sets without collecting more real data.
The residual generator is a drop-in replacement for the CGAN generator and can be combined with other discriminator regularizers.
The weight mask, derived purely from training samples and a template, is class-agnostic and could be applied to other conditional generation tasks.
The ablation implies the mask is the main driver of FID improvement, while residual units add a smaller gain; a practitioner might use the mask alone for distribution closeness.
The claimed detection and classification gains still need direct verification; if they hold, RWM-CGAN would be a practical augmentation tool for scarce-data domains.

Reading between the lines

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

A natural next test is to use the same augmentation pipeline in a few-shot classification benchmark, such as 5-way 1-shot, and compare test accuracy against standard augmentations; the paper does not run this experiment.
The difference-image mask is conceptually similar to saliency or anomaly maps; one could ablate whether the mask's benefit comes from suppressing background noise or from emphasizing class-discriminative pixels.
The mask is computed per class from that class's training samples; it would be worth testing whether a mask computed across all classes performs differently.
If the mask suppresses interference points, it may also regularize the discriminator against overfitting to idiosyncratic sample artifacts, which could be measured by comparing discriminator train and test accuracy.

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

You should know this one before spending time on it: the abstract promises “significant improvements in detection and classification accuracy” for few-shot learning, and the paper contains exactly zero detection or classification experiments. The only evidence is IS and FID on MNIST, and even that evidence is internally awkward: in Table III the full RWM-CGAN gets FID 10.126, worse than the weight-mask-only variant at 9.174. So the ablation does not support the claim that residual units help. The stress-test note is right on both counts.

What is actually new here is thin but not entirely absent. The authors combine two existing ideas — residual blocks in the generator and a weight mask in the discriminator — in a conditional GAN for few-shot data augmentation. That is a routine extension, not a conceptual contribution. To their credit, they do report per-class IS and FID numbers and an ablation table, and they are honest about class 7 being a failure case. That transparency counts for something, but it does not rescue the paper.

The soft spots are load-bearing. First, the headline claim about downstream accuracy is never tested. Better IS/FID on generated digits does not imply better few-shot classification or detection, and the paper makes no attempt to connect them. Second, the weight mask mechanism in Section III.B.2 is presented with corrupted equations — the inline math is unreadable, so the central innovation cannot be implemented or independently checked. Third, no code, no training details, no error bars or statistical tests; the reported differences could be noise. Fourth, the background section pads the reference list with irrelevant papers on stock prediction and medical forecasting, which suggests the authors did not engage deeply with the few-shot learning literature. The citations to He et al. and Mallya et al. are appropriate, so this is not a citation-integrity problem, just sloppiness.

My take: this is not ready for serious refereeing. The central empirical claim is unsupported, the method description is broken, and the ablation undercuts the core mechanism. A serious editor should desk reject it. If the authors return with actual few-shot accuracy experiments, a correct equation block, and code, it might be worth a look, but as it stands the paper is a weak engineering report on MNIST, not a contribution to few-shot learning.

Recommendation: desk reject.

Referee Report

5 major / 5 minor

Summary. The paper proposes RWM-CGAN, a conditional GAN that adds residual units to the generator and a weight-mask regularization to the discriminator, intended for few-shot learning data augmentation. The authors claim the method improves sample quality and, consequently, detection and classification accuracy. The experiments report Inception Score (IS) and Fréchet Inception Distance (FID) for MNIST digit generation, comparing CGAN and RWM-CGAN, plus an ablation study. The central claims about downstream task improvements are not tested, and the experimental design does not include a few-shot protocol.

Significance. If the method were validated, augmenting few-shot training sets with a controllable GAN would be a useful contribution to low-data regimes. The paper has a sensible high-level motivation and uses standard quantitative metrics (IS, FID), but the evidence presented is far too thin to support the stated claims. There are no downstream classification or detection experiments, no few-shot evaluation, no statistical significance measures, and no code or complete equations for the core mechanism. The ablation results contain an internal inconsistency that undermines the benefit of one of the two proposed components.

major comments (5)

[Abstract and Section V] The paper's headline claim—'significant improvements in detection and classification accuracy on public datasets'—is never evaluated. Section IV reports only IS and FID for MNIST digit generation; no detection or classification experiments appear anywhere. Since the entire motivation is few-shot learning via data augmentation, the absence of any downstream task evaluation leaves the central contribution empirically unsupported.
[Section IV.B, Table III] The ablation study is internally inconsistent with the paper's own justification. The full model (baseline+RU+WM) achieves FID 10.126, which is worse than the weight-mask-only variant (baseline+WM) at FID 9.174. The paper does not discuss this degradation, even though FID is the primary metric used to argue the method is superior. This result directly weakens the claim that residual units improve sample quality.
[Section III.B.2] The weight-mask equations are corrupted and unreadable; the displayed formulas contain missing symbols (e.g., '݀ ൌ ͳ ܯ෍ ெ ௠ୀଵ ...'), and the notation 'M' is not clearly defined. Because the weight mask is the central novel component, the method cannot be implemented, reproduced, or independently checked from the manuscript. This is a load-bearing omission, not a minor formatting issue.
[Section IV.A] The experimental setup does not match the few-shot learning setting that the paper claims to address. The MNIST dataset is used in its full form (60,000 training images), with no few-shot sampling protocol, no k-shot episodes, and no small-sample category constructions. Thus the experiments do not demonstrate effectiveness in the few-shot regime that motivates the work.
[Tables I and II] IS and FID are reported as single point estimates without error bars, standard deviations, or statistical tests. The text states that 2000 samples were evaluated 10 times on average, but no variance information is given. Without this, the reported differences (e.g., mean IS 6.415 vs. 6.966) cannot be assessed for significance, and the claim of consistent per-class improvement is not statistically supported.

minor comments (5)

[Reference [18]] Reference [18] cites 'Robust conditional generative adversarial networks' rather than the original CGAN paper by Mirza and Osindero; the original work should be cited for the CGAN baseline.
[Section III.B.2] The sentence 'The weight mask primarily involves two parameters: ݀ ሻand ሻrepresents...' has incomplete notation and undefined symbols; the two parameters should be explicitly named and their roles explained.
[Figure 3] Figure 3 is referenced in Section III.B.2 as the flow chart of the weight mask principle, but the figure content is not visible in the manuscript, and the surrounding text refers to missing elements.
[Table III] The caption says 'IS and FID results' but each row contains only one IS and one FID value; it should state that these are mean values over classes or over evaluation runs, along with any variance measures.
[Section I] The introduction states that 'extensive experiments demonstrate... significant improvements in both sample generation and downstream tasks,' but as noted in the major comments, no downstream experiments are reported; the wording should be revised to match the actual experimental content.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported IS/FID comparisons are empirical and not forced by construction, though the headline detection/classification accuracy claim is unsupported.

full rationale

This paper is an empirical data-augmentation study, not a derivation-driven one, so the classical circularity failure modes do not apply. The central reported results are Inception Score (IS) and Fréchet Inception Distance (FID) computed on MNIST-generated samples for baseline CGAN versus RWM-CGAN, and for the ablations in Table III. Nothing in the method section defines the evaluation metric in terms of the weight mask or vice versa: the mask is a data-dependent regularizer computed from training-sample difference images, and the metrics are computed on generated samples. That the mask is derived from the same training distribution used for evaluation is a design/validation concern rather than a circularity, because the reported numbers are not logically forced by the mask construction. The paper does invoke no uniqueness theorem, no fitted parameter renamed as a prediction, and no load-bearing self-citation chain; the few overlapping-author references are background citations and do not carry the argument. The abstract's claim of 'significant improvements in detection and classification accuracy' is not backed by any such experiments in Section IV, and Table III shows baseline+WM has a better FID (9.174) than the full baseline+RU+WM (10.126), but these are issues of empirical support and internal consistency, not circular reasoning. The corrupted equations in Section III.B.2 further impede reproducibility, but unreadable notation is not circularity. Accordingly, the appropriate circularity finding is a non-finding with score 0.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new theoretical entities. Its central mechanism, the weight mask, is a heuristic pipeline with unspecified parameters and no independent evidence. The assumptions are standard CGAN training plus an unvalidated mask heuristic and the implicit assumption that image quality metrics transfer to few-shot performance.

free parameters (2)

weight mask averaging count M
The weight mask is computed as an average over M difference images (Section III.B.2); M is not specified and appears to be a hand-chosen hyperparameter.
mask probability (dropout-like)
The paper states some weights are zeroed 'with a certain probability' but gives no value, leaving a key mechanism underspecified.

assumptions (4)

domain assumption Adversarial training of CGANs converges to a useful generator and discriminator on MNIST with the described architecture.
The paper provides no training curves or convergence analysis; it assumes standard CGAN training behaves as expected.
domain assumption Residual blocks improve generator sample quality in GANs.
Borrowed from image recognition literature (He et al.), assumed to transfer to GAN generators without direct justification.
ad hoc to paper Weight mask averaging over difference images suppresses interfering pixels and enhances small-sample features.
The mechanism is described qualitatively with garbled equations and no validation; this is the paper's core contribution and it is unsubstantiated.
domain assumption Inception Score and FID are appropriate proxies for data augmentation effectiveness in few-shot learning.
The paper uses IS/FID as the only evaluation, but does not connect these to downstream task accuracy.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Few-Shot Learning with Adaptive Weight Masking in Conditional GANs." pith.science (2026). https://pith.science/paper/HF7M5UAM

@misc{pith2026241203105,
  author       = {Pith},
  title        = {Pith review of: Few-Shot Learning with Adaptive Weight Masking in Conditional GANs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HF7M5UAM}},
  note         = {Machine review of arXiv:2412.03105}
}

read the original abstract

Deep learning has revolutionized various fields, yet its efficacy is hindered by overfitting and the requirement of extensive annotated data, particularly in few-shot learning scenarios where limited samples are available. This paper introduces a novel approach to few-shot learning by employing a Residual Weight Masking Conditional Generative Adversarial Network (RWM-CGAN) for data augmentation. The proposed model integrates residual units within the generator to enhance network depth and sample quality, coupled with a weight mask regularization technique in the discriminator to improve feature learning from small-sample categories. This method addresses the core issues of robustness and generalization in few-shot learning by providing a controlled and clear augmentation of the sample space. Extensive experiments demonstrate that RWM-CGAN not only expands the sample space effectively but also enriches the diversity and quality of generated samples, leading to significant improvements in detection and classification accuracy on public datasets. The paper contributes to the advancement of few-shot learning by offering a practical solution to the challenges posed by data scarcity and the need for rapid generalization to new tasks or categories.

Figures

Figures reproduced from arXiv: 2412.03105 by the authors.

**Figure 2.** Generator Residual block This approach introduces residual blocks into the generator to optimize the network architecture design. Each residual block comprises two convolutional layers with kernel sizes and strides of 3 and 1, respectively. To achieve identity mapping, ReLU activation functions and skip connections are employed between the convolutional layers. This network structure design effectively increases net… view at source ↗

**Figure 3.** Flow chart of weight mask principle y [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Discussion (0). Continue with ORCID 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

Dynamic Adaptation of LoRA Fine-Tuning for Efficient and Task-Specific Optimization of Large Language Models
cs.CL 2025-01 reject novelty 2.0 of 10

Dynamic LoRA, a layer-wise adaptive variant of LoRA, reportedly improves GLUE accuracy from 87.4% to 88.1% at only 0.1% more trainable parameters, but the write-up lacks reproducibility.

Reference graph

Works this paper leans on

22 extracted references · 19 canonical work pages · cited by 1 Pith paper

[1]

Remote sensing object detection in the deep learning era—a review[J]

Gui S, Song S, Qin R, et al. Remote sensing object detection in the deep learning era—a review[J]. Remote Sensing, 2024, 16(2): 327

work page 2024
[2]

Methods and datasets on semantic segmentation for Unmanned Aerial Vehicle remote sensing images: A review[J]

Cheng J, Deng C, Su Y, et al. Methods and datasets on semantic segmentation for Unmanned Aerial Vehicle remote sensing images: A review[J]. ISPRS Journal of Photog rammetry and Remote Sensing, 2024, 211: 1-34

work page 2024
[3]

Rsmamba: Remote sensing image classification with state space m odel[J]

Chen K, Chen B, Liu C, et al. Rsmamba: Remote sensing image classification with state space m odel[J]. IEEE Geoscience and Remote Sensing Letters, 2024

work page 2024
[4]

Text mining and natural language processing in construction[J]

Shamshiri A, Ryu K R, Park J Y. Text mining and natural language processing in construction[J]. Automation in Construction, 2024, 158: 105200

work page 2024
[5]

One-shot learning of object categories[J]

Fei-Fei L, Fergus R, Perona P. One-shot learning of object categories[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2006,28(4):594-611

work page 2006
[6]

Calibration Learning for Few-shot Novel Product Description,

Z. Liu, M. Wu, B. Peng, Y. Liu, Q. Peng, and C. Zou, "Calibration Learning for Few-shot Novel Product Description," in Proc. 46th Int. ACM SIGIR Conf. Res. Develop. Information Retrieval, pp. 1864- 1868, July 2023

work page 2023
[7]

Attention-Unet: A Deep Learning Approach for Fast and Accurate Segmentation in Medical Imaging,

Z. Zhu, Y. Yan, R. Xu, Y. Zi, and J. Wang, "Attention-Unet: A Deep Learning Approach for Fast and Accurate Segmentation in Medical Imaging," J. Comput. Sci. Softw. Appl., vol. 2, no. 4, pp. 24-31, 2022

work page 2022
[8]

Convolutional Neural Network Classification of Cancer Cytopathology Images: Taking Breast Cancer as an Example,

M. Xiao, Y. Li, X. Yan, M. Ga o, and W. Wang, "Convolutional Neural Network Classification of Cancer Cytopathology Images: Taking Breast Cancer as an Example," in Proc. 2024 7th Int. Conf. Mach. Vision Appl., pp. 145-149, Mar. 2024

work page 2024

Show all 22 references

[9]

Feature Extraction and Model Optimization of Deep Learning in Stock Market Prediction,

Y. Wei, X. Gu, Z. Fe ng, Z. Li, and M. Sun, "Feature Extraction and Model Optimization of Deep Learning in Stock Market Prediction," J. Comput. Technol. Softw., vol. 3, no. 4, 2024

2024
[10]

Predicting Stock Market Trends Using LSTM Networks: Overcoming RNN Limitations for Improved Financial Forecasting,

J. Wang, S. Hong, Y. Dong, Z. Li, and J. Hu, "Predicting Stock Market Trends Using LSTM Networks: Overcoming RNN Limitations for Improved Financial Forecasting," J. Comput. Sci. Softw. Appl., vol. 4, no. 3, pp. 1-7, 2024

2024
[11]

Adaptive Friction in Deep Learning: Enhancing Optimizers with Sigmoid and Tanh Function,

H. Zheng, B. Wang, M. Xiao, H. Qin, Z. Wu, and L. Tan, "Adaptive Friction in Deep Learning: Enhancing Optimizers with Sigmoid and Tanh Function," arXiv preprint arXiv:2408.11839, 2024

2024 arXiv
[12]

Dynamic Hypergraph-Enhanced Prediction of Sequential Medical Visits,

W. Yang, Z. Wu, Z. Zheng, B. Zhan g, S. Bo, and Y. Yang, "Dynamic Hypergraph-Enhanced Prediction of Sequential Medical Visits," arXiv preprint arXiv:2408.07084, 2024

2024 arXiv
[13]

Survival Prediction Across Diverse Cancer Types Using Neural Networks,

X. Yan, W. Wang, M. Xiao, Y. Li, and M. Gao, "Survival Prediction Across Diverse Cancer Types Using Neural Networks," in Proc. 2024 7th Int. Conf. Mach. Vision Appl., pp. 134-138, Mar. 2024

2024
[14]

Ndc- scene: Boost Monocular 3d Semantic Scene Completion in Normalized Device Coordinates Space,

J. Yao, C. Li, K. Sun, Y. Cai, H. Li, W. Ouyang, and H. Li, "Ndc- scene: Boost Monocular 3d Semantic Scene Completion in Normalized Device Coordinates Space," in Proc. 2023 IEEE/CVF Int. Conf. Comput. Vision (ICCV), pp. 9421-9431, Oct. 2023

2023
[15]

Exploring Anomaly Detection and Risk Assessment in Financial Markets Using Deep Neural Networks,

B. Wang, Y. Dong, J. Yao, H. Qin, and J. Wang, "Exploring Anomaly Detection and Risk Assessment in Financial Markets Using Deep Neural Networks," Int. J. Innov. Re s. Comput. Sci. Technol., vol. 12, no. 4, pp. 92-98, 2024

2024
[16]

Text Classification Optimization Algorithm Based on Graph Neural Network,

E. Gao, H. Yang, D. Sun, H. Xia, Y. Ma, and Y. Zhu, "Text Classification Optimization Algorithm Based on Graph Neural Network," arXiv preprint arXiv:2408.15257, 2024

2024 arXiv
[17]

Advancing Emotional Analysis with Large Language Models,

H. Yang, Y. Zi, H. Qin, H. Zhen g, and Y. Hu, "Advancing Emotional Analysis with Large Language Models," J. Comput. Sci. Softw. Appl., vol. 4, no. 3, pp. 8-15, 2024

2024
[18]

Robust conditional generative adversarial networks[J]

Chrysos G G, Kossaifi J, Zafeiriou S. Robust conditional generative adversarial networks[J]. arXiv preprint arXiv:1805.08657, 2018

2018 arXiv
[19]

Deep residual learning for image recognition[C]//Proceedings of the IEEE conference on computer vision and pattern recognition

He K, Zhang X, Ren S, et al. Deep residual learning for image recognition[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2016: 770-778

2016
[20]

Piggyback: Adapting a single network to multiple tasks by learning to mask weights[C]//Proceedings of the European conference on computer vision (ECCV)

Mallya A, Davis D, Lazebnik S. Piggyback: Adapting a single network to multiple tasks by learning to mask weights[C]//Proceedings of the European conference on computer vision (ECCV). 2018: 67-82

2018
[21]

(2023, December)

Li, Y., Yan, X., Xiao, M., Wang, W ., & Zhang, F. (2023, December). Investigation of creating accessibility linked data based on publicly available accessibility datasets. In Proceedings of the 2023 13th International Conference on Communication and Network Security, (pp. 77-81)

2023
[22]

Barratt, S., & Sharma, R. (2018). A note on the inception score. arXiv preprint arXiv:1801.01973

2018 arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [1]

Remote sensing object detection in the deep learning era—a review[J]

Gui S, Song S, Qin R, et al. Remote sensing object detection in the deep learning era—a review[J]. Remote Sensing, 2024, 16(2): 327

work page 2024

[2] [2]

Methods and datasets on semantic segmentation for Unmanned Aerial Vehicle remote sensing images: A review[J]

Cheng J, Deng C, Su Y, et al. Methods and datasets on semantic segmentation for Unmanned Aerial Vehicle remote sensing images: A review[J]. ISPRS Journal of Photog rammetry and Remote Sensing, 2024, 211: 1-34

work page 2024

[3] [3]

Rsmamba: Remote sensing image classification with state space m odel[J]

Chen K, Chen B, Liu C, et al. Rsmamba: Remote sensing image classification with state space m odel[J]. IEEE Geoscience and Remote Sensing Letters, 2024

work page 2024

[4] [4]

Text mining and natural language processing in construction[J]

Shamshiri A, Ryu K R, Park J Y. Text mining and natural language processing in construction[J]. Automation in Construction, 2024, 158: 105200

work page 2024

[5] [5]

One-shot learning of object categories[J]

Fei-Fei L, Fergus R, Perona P. One-shot learning of object categories[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2006,28(4):594-611

work page 2006

[6] [6]

Calibration Learning for Few-shot Novel Product Description,

Z. Liu, M. Wu, B. Peng, Y. Liu, Q. Peng, and C. Zou, "Calibration Learning for Few-shot Novel Product Description," in Proc. 46th Int. ACM SIGIR Conf. Res. Develop. Information Retrieval, pp. 1864- 1868, July 2023

work page 2023

[7] [7]

Attention-Unet: A Deep Learning Approach for Fast and Accurate Segmentation in Medical Imaging,

Z. Zhu, Y. Yan, R. Xu, Y. Zi, and J. Wang, "Attention-Unet: A Deep Learning Approach for Fast and Accurate Segmentation in Medical Imaging," J. Comput. Sci. Softw. Appl., vol. 2, no. 4, pp. 24-31, 2022

work page 2022

[8] [8]

Convolutional Neural Network Classification of Cancer Cytopathology Images: Taking Breast Cancer as an Example,

M. Xiao, Y. Li, X. Yan, M. Ga o, and W. Wang, "Convolutional Neural Network Classification of Cancer Cytopathology Images: Taking Breast Cancer as an Example," in Proc. 2024 7th Int. Conf. Mach. Vision Appl., pp. 145-149, Mar. 2024

work page 2024