Scholarlink

In this research, we developed a Generative Adversarial Network (GAN) to create synthetic images of chest X-rays for detecting normal and pneumonia cases. We started with an exploratory data analysis of the Chest X-ray Pneumonia dataset, which included 1341 normal images and 3875 pneumonia images. To enhance our dataset, we incorporated additional images from the NIH ChestX-ray14 and Chest Xpert datasets, expanding the total to 7500 images—3000 normal and 4500 pneumonias. With the framework of Tensor Flow, we trained and constructed our GAN with multi-GPU support. As the GAN model has two neurons, known as the generator and the discriminator. Where in This model included several Conv2DTranspose and Conv2D layers to manage image generation and image classification. The GAN model is trained for 60 epochs, during which we performed ND and monitored various evaluation metrics. In first stage we got the discriminator loss was 0.4859, and the generator loss was 1.2401. By the end of the stage in training, the discriminator loss stabilized at 0.6303, and the generator loss settled at around 0.8465, indicating effective training dynamics. Our GAN achieved a Fréchet Inception Distance (FID) score of 18.5, which is better than the scores of Conditional GANs (cGANs) and StyleGANs, recorded at 22.7 and 25.3, respectively. We also calculated the Structural Similarity Index (SSIM), with our model attaining an average score of 0.89. Additionally, when we integrated the synthetic images into a classifier, we observed an improvement in classification accuracy from 85 to 90%. These results show that our GAN can produce realistic chest X-ray images, providing a valuable resource for data augmentation in medical imaging.

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

Generating Synthetic Chest X-ray Images for Pneumonia Detection Using Generative Adversarial Networks

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Vaduguru Venkata Ramya,
  • Anju Khandelwal

摘要

In this research, we developed a Generative Adversarial Network (GAN) to create synthetic images of chest X-rays for detecting normal and pneumonia cases. We started with an exploratory data analysis of the Chest X-ray Pneumonia dataset, which included 1341 normal images and 3875 pneumonia images. To enhance our dataset, we incorporated additional images from the NIH ChestX-ray14 and Chest Xpert datasets, expanding the total to 7500 images—3000 normal and 4500 pneumonias. With the framework of Tensor Flow, we trained and constructed our GAN with multi-GPU support. As the GAN model has two neurons, known as the generator and the discriminator. Where in This model included several Conv2DTranspose and Conv2D layers to manage image generation and image classification. The GAN model is trained for 60 epochs, during which we performed ND and monitored various evaluation metrics. In first stage we got the discriminator loss was 0.4859, and the generator loss was 1.2401. By the end of the stage in training, the discriminator loss stabilized at 0.6303, and the generator loss settled at around 0.8465, indicating effective training dynamics. Our GAN achieved a Fréchet Inception Distance (FID) score of 18.5, which is better than the scores of Conditional GANs (cGANs) and StyleGANs, recorded at 22.7 and 25.3, respectively. We also calculated the Structural Similarity Index (SSIM), with our model attaining an average score of 0.89. Additionally, when we integrated the synthetic images into a classifier, we observed an improvement in classification accuracy from 85 to 90%. These results show that our GAN can produce realistic chest X-ray images, providing a valuable resource for data augmentation in medical imaging.