基于自注意力机制生成对抗网络的医学图像生成

Medical image generation based on improved generative adversarial

针对图像生成算法中生成对抗网络训练效率低且不稳定和原始乳腺癌数据集分布不均匀等问题,提出一种改进的SAGAN模型,在生成图像任务中表现更好,相较传统SAGAN、GAN、DCGAN模型,它的关键改进是使用ReLU6激活函数和铰链损失函数,取代了原有的ReLU激活函数和二分类平衡交叉熵损失函数,这些改进提高了生成图像的质量、多样性和训练稳定性。实验结果表明,改进的SAGAN的D-Loss相较传统SAGAN下降了0.114,均方误差(MSE)下降了0.09,结构相似性(SSIM)增加了0.04。说明改进的SAGAN在生成高质量图像和更好地保留图像结构方面具有优势。

Aiming at the problems of low and unstable training efficiency of generative adversarial networks and uneven distribution of the original breast cancer dataset in image generation algorithms,this paper proposes an improved SAGAN model,which performs better in the task of generating images,and compared with the traditional SAGAN and GAN,DCGAN models,its key improvement is the use of the ReLU6 activation function and the hinge loss function,instead of the original ReLU activation function and binary balanced cross-entropy loss function,and these improvements improve the quality,diversity and training stability of the generated images.The experimental results show that the D-Loss of the improved SAGAN decreases by 0.114,the mean square error(MSE)decreases by 0.09,and the structural similarity(SSIM)increases by 0.04 compared tOthe conventional SAGAN.This indicates that the improved SAGAN has an advantage in generating high-quality images and better preserving the image structure.