基于迁移学习的糖尿病视网膜病变眼底图像合成研究

Image Synthesis of Diabetic Retinopathy Based on Transfer Learning: A Study

摘要: 文章研究了利用生成对抗网络(GAN)和迁移学习在小数据集上生成糖尿病视网膜病变(DR)眼底图像的有效方法。传统的图像生成技术在高维数据下表现不佳，难以准确学习数据分布。自2014年GAN被提出以来，众多研究致力于其优化与应用，然而在小样本情况下仍面临训练难度和过拟合问题。文章通过迁移学习方法，利用在大数据集上预训练的GAN模型，结合条件生成对抗网络(CGAN)，在小数据集上重新训练部分参数，从而生成高清晰度的DR图像，并提高生成图像的类别控制能力。研究表明，采用迁移学习的方法显著提升了图像生成质量，实验结果显示，基于迁移学习的GAN和CGAN在小数据集上的FID分数明显优于无迁移学习的对照组，表明生成的图像在质量和多样性上有显著改善。这项研究不仅为医学图像生成提供了新的思路，也为解决小数据集和不平衡问题提供了有效的解决方案。

Abstract: This paper investigates an effective method for generating fundus images of diabetic retinopathy (DR) on small datasets using Generative Adversarial Networks (GANs) and transfer learning. Traditional image generation techniques struggle to accurately learn data distributions in high-dimensional data, leading to poor performance. Since the introduction of GANs in 2014, numerous studies have focused on their optimization and applications; however, they still face challenges such as training difficulties and overfitting in small sample cases. This study employs transfer learning by utilizing a GAN model pre-trained on a large dataset, combined with Conditional Generative Adversarial Networks (CGAN), to retrain some parameters on a small dataset. This approach generates high-quality DR images and enhances the class control capability of the generated images. Research shows that using transfer learning significantly improves the quality of image generation. Experimental results indicate that the FID scores of GANs and CGANs based on transfer learning are notably superior to those of the control group without transfer learning, demonstrating significant improvements in the quality and diversity of the generated images. This research not only provides new insights for medical image generation but also offers effective solutions for addressing small datasets and imbalance issues.