pith. sign in

arxiv: 2605.01815 · v2 · pith:CDVOAG3Znew · submitted 2026-05-03 · 💻 cs.CV

Cross-Domain Adversarial Augmentation: Stabilizing GANs for Medical and Handwriting Data Scarcity

Md. Sohanuzzaman Soad , Mahady Al Hady , S M Rafiuddin Rifat , Sudip Ghose This is my paper

Pith reviewed 2026-05-21 00:15 UTC · model grok-4.3

classification 💻 cs.CV
keywords GAN augmentationdata scarcityBangla handwriting recognitionchest X-ray analysissynthetic dataimage classificationDCGANlow-resource domains
0
0 comments X

The pith

Adding GAN-generated synthetic images to small training sets improves classification accuracy for Bangla handwriting and chest X-rays.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tests whether GANs can produce synthetic training images that help overcome data shortages in specialized computer vision tasks. It applies this to two low-resource cases: recognizing Bangla handwritten characters and classifying chest X-ray images. Classifiers trained on mixtures of real and synthetic samples achieve higher performance than those using real data alone, with supporting checks on image quality and training stability. The approach matters because many practical imaging problems lack enough labeled examples for reliable deep learning models. The authors also note practical hurdles such as privacy in medical contexts and the need for careful evaluation of the generated samples.

Core claim

DCGAN-based models generate synthetic 64x64 samples that, when combined with limited real data, increase training diversity and raise downstream classification accuracy in Bangla handwritten character recognition and chest X-ray analysis. Quality is assessed via Inception Score, Fréchet Inception Distance, t-SNE, and UMAP visualizations, while ablation studies examine synthetic-to-real ratios, sample filtering, and stability methods such as gradient penalty and spectral normalization.

What carries the argument

DCGAN generative augmentation that produces synthetic images to supplement scarce real data, then feeds the mixed set into image classifiers to test performance gains.

If this is right

  • Synthetic augmentation raises classifier accuracy in limited-data settings for the two domains examined.
  • Gradient penalty and spectral normalization help stabilize GAN training for these image types.
  • Ablation experiments identify useful synthetic-to-real ratios and filtering strategies.
  • Synthetic data offers a route to address scarcity while raising questions about medical image evaluation and privacy.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same augmentation pattern may help other low-resource visual domains such as rare disease detection or specialized satellite imagery.
  • Combining this with domain-adaptation methods could reduce risks from distribution shifts between real and generated images.
  • Privacy-preserving synthetic generation could become routine in regulated fields where real data sharing is restricted.

Load-bearing premise

Mixing GAN-generated samples into the training set will improve results on unseen real test images without the synthetics introducing artifacts or shifts that hurt accuracy on actual data.

What would settle it

Compare accuracy of a classifier trained only on real data against one trained on real plus synthetic data when both are tested on the same held-out set of real images; if the mixed version does not outperform or underperforms, the augmentation benefit is not shown.

Figures

Figures reproduced from arXiv: 2605.01815 by Mahady Al Hady, Md. Sohanuzzaman Soad, S M Rafiuddin Rifat, Sudip Ghose.

Figure 1
Figure 1. Figure 1: Overview of proposed solution 3.1 Dataset Description BanglaLekha Isolated, many handwritten Bangla characters that comprise numerals, simple letters, and compound forms, is the initial dataset used in this investigation. It includes over 160,000 samples taken from individuals in various parts of Bangladesh and at various ages [1]. This dataset is particularly helpful for testing how effectively GAN models… view at source ↗
Figure 2
Figure 2. Figure 2: Sample Images of the BanglaLekha Isolated Dataset [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Sample images from the COVID-19 Chest X-Ray Dataset, illustrating variability across patient cases and [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Example images from the preprocessed BanglaLekha Isolated dataset. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visual assessment of GAN-generated samples across the BanglaLekha Isolated and COVID-19 Chest X-ray [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Original Generative Adversarial Network Training Algorithm Source: Generative Adversarial Networks [3]. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 5
Figure 5. Figure 5: Training behavior of the DCGAN model, showing generator–discriminator loss dynamics and evaluation [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Original Algorithm of t-SNE Source: Visualizing Data using t-SNE. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 6
Figure 6. Figure 6: DCGAN Real and Fake Score of Generated Image During Training on the Bangla Lekha Isolated dataset. [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Images generated of Bangla Numeric Characters by GAN after 200 epochs of training. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 7
Figure 7. Figure 7: Embedding-space visualization of real and GAN-generated Bangla numeric characters using two dimen [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Generated Images of Bangla Numeric Characters by GAN at different stages of training. [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Generated Sample from COVID-19 Chest X-ray Dataset During Training. [PITH_FULL_IMAGE:figures/full_fig_p009_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Final generated individual character of the Bangla Lekha Isolated dataset [PITH_FULL_IMAGE:figures/full_fig_p009_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: DCGAN training loss for discriminators and generator on the Bangla Lekha Isolated dataset. [PITH_FULL_IMAGE:figures/full_fig_p010_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: DCGAN Real and Fake Score of Generated Image During Training on the Bangla Lekha Isolated dataset. [PITH_FULL_IMAGE:figures/full_fig_p010_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Evaluation Metric During Training of DCGAN [PITH_FULL_IMAGE:figures/full_fig_p011_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: PCA 2-D Plot of Generated Bangla Numeric Character. [PITH_FULL_IMAGE:figures/full_fig_p011_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: t-SNE 2-D Plot of Generated Bangla Numeric Character. [PITH_FULL_IMAGE:figures/full_fig_p012_15.png] view at source ↗
read the original abstract

Generative Adversarial Networks (GANs) can help overcome data scarcity in computer vision tasks by generating additional training samples. In this work, we explore generative data augmentation in two low-resource domains: Bangla handwritten character recognition and chest X-ray image analysis. We use DCGAN-based models trained on 64x64 images to generate synthetic samples and evaluate their quality using Inception Score (IS), Fr\'echet Inception Distance (FID), and visualization methods such as t-SNE and UMAP. To measure practical usefulness, we train image classifiers using real data and a combination of real and synthetic data. Experimental results show that synthetic augmentation improves data diversity and consistently increases classification performance in limited-data settings. We also investigate training stability techniques, including gradient penalty and spectral normalization, and perform ablation studies on synthetic-to-real data ratios and sample filtering strategies. In addition, we discuss challenges related to medical image evaluation, dataset licensing, and privacy concerns of synthetic data. Our approach is simple, reproducible, and provides a strong baseline for generative augmentation in resource-constrained imaging applications.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript explores DCGAN-based generative augmentation to address data scarcity in two domains: Bangla handwritten character recognition and chest X-ray analysis. DCGANs are trained on 64x64 images to produce synthetic samples, which are evaluated via Inception Score, Fréchet Inception Distance, t-SNE, and UMAP. Classifiers are then trained on real-only versus real-plus-synthetic mixtures; the authors report that augmentation improves diversity and yields consistent gains in classification accuracy under limited-data regimes. Additional contributions include stability techniques (gradient penalty, spectral normalization), ablations on synthetic-to-real ratios and filtering, and discussion of medical-image evaluation, licensing, and privacy issues.

Significance. If the empirical gains are shown to be robust and to generalize to strictly held-out real test images, the work supplies a simple, reproducible baseline for generative augmentation in resource-constrained medical and document-image settings. The emphasis on downstream classifier accuracy rather than GAN metrics alone, together with explicit ablations and domain-specific caveats, increases practical utility. The absence of quantitative numbers, error bars, or statistical tests in the current presentation, however, prevents a full appraisal of effect size and reliability.

major comments (3)
  1. [Abstract and §4] Abstract and §4 (Experimental Results): the central claim that 'synthetic augmentation ... consistently increases classification performance' is asserted without any reported accuracy values, dataset cardinalities, baseline numbers, or error bars. This omission renders the magnitude and reliability of the reported gains impossible to assess from the manuscript as written.
  2. [§3.2 and §4.1] §3.2 and §4.1 (Evaluation Protocol): it is not stated whether the test partition consists exclusively of real images that were never used to train either the DCGAN or the downstream classifier. In low-data regimes this distinction is load-bearing for the claim that observed gains reflect useful diversity rather than distribution shift or leakage; an explicit protocol description and, ideally, a statement that test images are strictly external real samples are required.
  3. [§4.3] §4.3 (Ablations): while ratios and filtering strategies are examined, no statistical significance tests (e.g., paired t-tests across random seeds) or variance estimates accompany the performance curves. Without these, it is difficult to determine whether the reported improvements are robust or could be explained by training stochasticity.
minor comments (2)
  1. [Figures 3-5] Figure captions for t-SNE/UMAP embeddings should explicitly note the number of real versus synthetic points plotted and the perplexity or nearest-neighbor parameters used.
  2. [Table 1] The manuscript would benefit from a short table summarizing the exact number of real training images available in each limited-data regime before and after augmentation.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We agree that the presentation of results requires additional quantitative detail and explicit protocol clarifications to allow proper assessment of the claims. We address each major comment below and will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [Abstract and §4] Abstract and §4 (Experimental Results): the central claim that 'synthetic augmentation ... consistently increases classification performance' is asserted without any reported accuracy values, dataset cardinalities, baseline numbers, or error bars. This omission renders the magnitude and reliability of the reported gains impossible to assess from the manuscript as written.

    Authors: We acknowledge this omission in the current draft. While the full experimental section contains tables and figures showing accuracy improvements for real-only versus real-plus-synthetic training, specific numerical values, dataset sizes, and error bars were not highlighted in the abstract or summarized in §4. We will add a consolidated results table with mean accuracies, standard deviations across seeds, baseline comparisons, and dataset cardinalities to the abstract and §4. revision: yes

  2. Referee: [§3.2 and §4.1] §3.2 and §4.1 (Evaluation Protocol): it is not stated whether the test partition consists exclusively of real images that were never used to train either the DCGAN or the downstream classifier. In low-data regimes this distinction is load-bearing for the claim that observed gains reflect useful diversity rather than distribution shift or leakage; an explicit protocol description and, ideally, a statement that test images are strictly external real samples are required.

    Authors: The evaluation protocol in §3.2 uses a strict hold-out of real images for testing that are excluded from both DCGAN training and classifier training. However, we agree the description is insufficiently explicit. We will revise §3.2 and §4.1 to state clearly that all test images are real samples never seen during GAN or classifier training, and we will add a diagram or pseudocode of the data split to remove any ambiguity. revision: yes

  3. Referee: [§4.3] §4.3 (Ablations): while ratios and filtering strategies are examined, no statistical significance tests (e.g., paired t-tests across random seeds) or variance estimates accompany the performance curves. Without these, it is difficult to determine whether the reported improvements are robust or could be explained by training stochasticity.

    Authors: We agree that variance estimates and significance testing are necessary for robustness claims. The current ablations report single-run curves. We will re-run the key experiments across multiple random seeds, add error bars to the performance plots, and include paired t-test p-values comparing real-only versus augmented settings in the revised §4.3. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical results rely on external metrics

full rationale

The paper presents an empirical study of DCGAN-based augmentation for Bangla handwriting and chest X-ray classification. All load-bearing claims rest on direct measurements (IS, FID, t-SNE/UMAP visualizations, and downstream classifier accuracy) computed against standard external benchmarks and held-out real test partitions. No mathematical derivation, first-principles prediction, or fitted parameter is renamed as an independent result; the reported gains are simple before/after comparisons on real-plus-synthetic training sets. The work is therefore self-contained against external evaluation protocols and contains no self-definitional, fitted-input, or self-citation load-bearing steps.

Axiom & Free-Parameter Ledger

1 free parameters · 0 axioms · 0 invented entities

Experimental study relying on standard GAN training assumptions and off-the-shelf metrics; no new mathematical axioms or invented entities are introduced.

free parameters (1)
  • synthetic-to-real ratio
    Ablation studies vary this ratio to optimize performance, indicating it functions as a tuned hyperparameter.

pith-pipeline@v0.9.0 · 5740 in / 1079 out tokens · 57135 ms · 2026-05-21T00:15:11.051939+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

9 extracted references · 9 canonical work pages

  1. [1]

    Shopon, Nabeel Mohammed, Sifat Momen, and Md

    Mithun Biswas, Rafiqul Islam, Gautam Kumar Shom, Md. Shopon, Nabeel Mohammed, Sifat Momen, and Md. Anowarul Abedin. BanglaLekha-Isolated: A multi-purpose comprehensive dataset of handwritten bangla isolated characters.Data in Brief, 12:103–107, June 2017

    work page 2017

  2. [2]

    Towards robust sta- bility prediction in smart grids: GAN-based approach under data constraints and adversarial challenges.Internet of Things, 33:101662, September 2025

    Emad Efatinasab, Alessandro Brighente, Denis Donadel, Mauro Conti, and Mirco Rampazzo. Towards robust sta- bility prediction in smart grids: GAN-based approach under data constraints and adversarial challenges.Internet of Things, 33:101662, September 2025

    work page 2025

  3. [3]

    Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio

    Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. InAdvances in Neural Information Processing Sys- tems, volume 27, pages 2672–2680. Curran Associates, Inc., 2014

    work page 2014

  4. [4]

    Mehedi Hassan, Md

    Md. Mehedi Hassan, Md. Ashik Mahmud, Abrar Shahriyar, Naquibuddin Sarkar, Sonjoy Chandra Mohonto, Md Jakir Hossain, and Golam Rakib Chowdhury. Smart spectacles for the deaf with voice to text and sign language integration. In2023 26th International Conference on Computer and Information Technology (ICCIT), pages 2671–2676, Cox’s Bazar, Bangladesh, Decembe...

    work page 2023

  5. [5]

    Towards improved evaluation of generative neural networks: The Fr´echet coefficient.Neurocomputing, 623:129422, March 2025

    Adrian Kucharski and Anna Fabija ´nska. Towards improved evaluation of generative neural networks: The Fr´echet coefficient.Neurocomputing, 623:129422, March 2025. 10

    work page 2025

  6. [6]

    Future of generative adver- sarial networks (GAN) for anomaly detection in network security: A review.Computers & Security, 139:103733, April 2024

    Willone Lim, Kelvin Sheng Chek Yong, Bee Theng Lau, and Colin Choon Lin Tan. Future of generative adver- sarial networks (GAN) for anomaly detection in network security: A review.Computers & Security, 139:103733, April 2024

    work page 2024

  7. [7]

    Detection of COVID-19 from chest X-Ray images using convolutional neural networks.SLAS Technology: Translating Life Sciences Innovation, 25(6):553–565, September 2020

    Boran Sekeroglu and Ilker Ozsahin. Detection of COVID-19 from chest X-Ray images using convolutional neural networks.SLAS Technology: Translating Life Sciences Innovation, 25(6):553–565, September 2020

    work page 2020

  8. [8]

    Satvik Tripathi, Alisha Isabelle Augustin, Adam Dunlop, Rithvik Sukumaran, Suhani Dheer, Alex Zavalny, Owen Haslam, Thomas Austin, Jacob Donchez, Pushpendra Kumar Tripathi, and Edward Kim. Recent advances and ap- plication of generative adversarial networks in drug discovery, development, and targeting.Artificial Intelligence in the Life Sciences, 2:10004...

    work page 2022

  9. [9]

    Generative adversarial networks in medical image segmentation: A review.Computers in Biology and Medicine, 140:105063, January 2022

    Siyi Xun, Dengwang Li, Hui Zhu, Min Chen, Jianbo Wang, Jie Li, Meirong Chen, Bing Wu, Hua Zhang, Xiangfei Chai, Zekun Jiang, Yan Zhang, and Pu Huang. Generative adversarial networks in medical image segmentation: A review.Computers in Biology and Medicine, 140:105063, January 2022. 11

    work page 2022