The Impact of Gradient Penalty Parameter on the Performance of Deep Convolutional Generative Adversarial Network with Wasserstein Penalty in Mammography
Mammography is a significant tool in the detection of breast cancer in the initial phases, but automated diagnostic system development is hindered by the lack of annotated data and the cost of expert labeling. To overcome such limitations, a Deep Convolutional Generative Adversarial Network (DCGAN) with a Wasserstein Gradient Penalty (W-GP) was investigated to generate mammographic images. The primary goal here is to assess the effect of different values of the gradient penalty parameter (λ = 0.5, 3.0, and 5.0) on the generated images' quality, realism, and structural accuracy. The CBIS-DDSM dataset was used to focus on calcification neoplasms. The generated images were evaluated through KID, LPIPS, and feature-space metrics. Results demonstrate that higher λ values improve distributional alignment, while lower values enhance perceptual similarity. Moreover, λ should be carefully adjusted based on application needs with a balance between statistical accuracy and visual realism
Сведения
- Журнал
- Trudy Universiteta
- Опубликовано
- 2025 · Т. 4, № 101, С. 427
- Язык
- Английский
- ISSN
- 1609-1825 · 2710-3382
- Издатель
- Karaganda Technical University named after Abylkas Saginov
- Аффилиации
- Akhmet Baitursynuly Kostanay Regional University