Scholarlink

Generative AI for image-to-image generation delivers state-of-the-art results in various computer vision tasks, often requiring extensive datasets for accurate and high-resolution image synthesis. Current convolutional models like StyleGAN tend to focus on local patterns but neglect global patterns. StyleSwin, a transformer-based image generative model, has emerged as a strong competitor because it captures both, enhancing image quality. However, training StyleSwin with limited data can lead to discriminator overfitting, causing training instability. This issue is particularly relevant for medical data, which are often limited and multiclass. To address this, we propose an adaptive discriminator augmentation (ADA) mechanism for StyleSwin, aiming to stabilize training in data-constrained environments. The original StyleSwin implementation utilizes Swin Transformer blocks in the generator and included a style injection mechanism, achieving state-of-the-art results in high-resolution image generation. Our approach addresses medical imaging challenges such as the scarcity of high-resolution images, limited data, and artifacts. By incorporating ADA, we have generated high-resolution medical images despite data limitations. In this study, we segmented and cropped blood cell images to extract the main cells. We then trained StyleSwin with ADA on these pre-processed hematologic images (blood cells), successfully generating artifact-free multiclass images. To evaluate the patterns of the generated images, we used Fréchet Inception Distance (FID) score. Although StyleSwin-generated images achieved a lower FID score compared to those generated by StyleSwin-ADA, indicating an overall statistically significant difference (p = 0.016225) calculated by a paired t-test, FID focuses mainly on overall statistical similarity and may not capture finer details. Therefore, to gain a more comprehensive assessment, we additionally employed multivariate principal component analysis (PCA) to analyze the patterns within the generated images. The results suggest that StyleSwin-ADA provides a more distinct and clear representation compared to StyleSwin, indicating better preservation of the pattern of the real imaging data and improved noise reduction.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Training Swin Transformer-Based GAN for High-Resolution Blood Cell Imaging with Limited Data

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Tanviben Patel,
  • Hoda El-Sayed,
  • Md Kamruzzaman Sarker,
  • Dat Duong,
  • Benjamin D. Solomon

摘要

Generative AI for image-to-image generation delivers state-of-the-art results in various computer vision tasks, often requiring extensive datasets for accurate and high-resolution image synthesis. Current convolutional models like StyleGAN tend to focus on local patterns but neglect global patterns. StyleSwin, a transformer-based image generative model, has emerged as a strong competitor because it captures both, enhancing image quality. However, training StyleSwin with limited data can lead to discriminator overfitting, causing training instability. This issue is particularly relevant for medical data, which are often limited and multiclass. To address this, we propose an adaptive discriminator augmentation (ADA) mechanism for StyleSwin, aiming to stabilize training in data-constrained environments. The original StyleSwin implementation utilizes Swin Transformer blocks in the generator and included a style injection mechanism, achieving state-of-the-art results in high-resolution image generation. Our approach addresses medical imaging challenges such as the scarcity of high-resolution images, limited data, and artifacts. By incorporating ADA, we have generated high-resolution medical images despite data limitations. In this study, we segmented and cropped blood cell images to extract the main cells. We then trained StyleSwin with ADA on these pre-processed hematologic images (blood cells), successfully generating artifact-free multiclass images. To evaluate the patterns of the generated images, we used Fréchet Inception Distance (FID) score. Although StyleSwin-generated images achieved a lower FID score compared to those generated by StyleSwin-ADA, indicating an overall statistically significant difference (p = 0.016225) calculated by a paired t-test, FID focuses mainly on overall statistical similarity and may not capture finer details. Therefore, to gain a more comprehensive assessment, we additionally employed multivariate principal component analysis (PCA) to analyze the patterns within the generated images. The results suggest that StyleSwin-ADA provides a more distinct and clear representation compared to StyleSwin, indicating better preservation of the pattern of the real imaging data and improved noise reduction.