EUVIMEDEUVIMED entdecken
Crossref

Enhancing rare tumor detection: A cross-modal generative adversarial network benchmark for data augmentation in cardiac, liver, and retinal imaging

Muhammad Umer Farooq, Danish Jamil, Saad Bin Jawaid

Eurasian Journal of Medicine and Oncology · 30.06.2026 · 10 (3) · 025460482

DOI: 10.36922/ejmo025460482 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Eurasian Journal of Medicine and Oncology, online erschienen am 30.06.2026. Verfasst von Muhammad Umer Farooq und 2 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Introduction

Rare pathologies in medical imaging suffer from severe data scarcity, leading to AI models with low sensitivity and high false-negative rates, resulting in missed diagnoses with potentially life-threatening consequences. Although generative adversarial networks (GANs) offer a promising solution by generating synthetic images, no empirically derived quality thresholds currently exist for safe clinical deployment.

Objective

To systematically evaluate DCGAN, WGAN-GP, and StyleGAN for generating clinically useful synthetic chest X-rays, focusing on false-negative reduction and establishing clinical quality thresholds.

Methods

Using the NIH Chest X-ray dataset (89,139 images; 14 pathologies), three GAN architectures were trained to augment underrepresented classes. Evaluation included quantitative metrics (FID and structure-specific SSIM), diagnostic performance across three datasets, blinded radiologist review (n = 5; 100 images per model), and failure analysis quantifying false-negative rates.

Results

StyleGAN outperformed alternatives (FID = 18.2 vs. DCGAN: 45.6; WGAN-GP: 23.4), achieved SSIM of 0.92 (vs. 0.78 and 0.85), and preserved lung patterns at 0.90 (vs. 0.74 and 0.82). Sensitivity increased from 79.5% to 94.2%, yielding approximately 10 additional early detections per 100 rare pathology cases. StyleGAN reduced false negatives for small nodules to 12% compared to 28% for DCGAN—a 16% absolute reduction, translating to 160 additional correct diagnoses per 1,000 high-risk screenings. Radiologists rated StyleGAN images 4.7/5 (vs. DCGAN: 2.8/5). This study proposes the first empirically derived clinical quality thresholds for synthetic chest X-rays: FID < 20, SSIM > 0.90, small-structure SSIM > 0.85, and radiologist score > 4.5/5. Only StyleGAN met all criteria.

Conclusion

High-quality GANs, particularly StyleGAN, significantly reduce false negatives and improve rare pathology detection. By directly linking synthetic image quality to measurable reductions in false negatives, this study establishes clinically actionable safety thresholds and provides a regulatory-aligned framework for responsible deployment of GAN-augmented medical imaging systems.

Zitieren

APA

Farooq, M. U., Jamil, D., & Jawaid, S. B. (2026). Enhancing rare tumor detection: A cross-modal generative adversarial network benchmark for data augmentation in cardiac, liver, and retinal imaging. Eurasian Journal of Medicine and Oncology, 10(3), 025460482. https://doi.org/10.36922/ejmo025460482

Vancouver

Farooq MU, Jamil D, Jawaid SB. Enhancing rare tumor detection: A cross-modal generative adversarial network benchmark for data augmentation in cardiac, liver, and retinal imaging. Eurasian Journal of Medicine and Oncology. 2026;10(3):025460482. doi:10.36922/ejmo025460482

BibTeX
@article{Farooq2026,
  author = {Farooq, Muhammad Umer and Jamil, Danish and Jawaid, Saad Bin},
  title = {Enhancing rare tumor detection: A cross-modal generative adversarial network benchmark for data augmentation in cardiac, liver, and retinal imaging},
  journal = {Eurasian Journal of Medicine and Oncology},
  year = {2026},
  volume = {10},
  number = {3},
  pages = {025460482},
  doi = {10.36922/ejmo025460482}
}
Kostenlos starten →