摘要： 在医疗领域，许多疾病的诊断依赖高倍数显微镜对细胞等微观物体的观测，但由于高倍数显微镜价格昂贵，操作复杂，且高倍数细胞显微图像重建工作存在低、高倍数显微图像之间图片风格不统一、细胞图像清晰度不一致和训练数据不匹配等问题。为此，提出高倍数细胞显微图像生成式对抗网络。将全新激活函数引入CycleGAN网络，在生成器中添加新的残差密集块并去掉BN层。同时为确保生成图像真实可信，在生成器训练过程中考虑细节感知损失。实验结果表明，该方法在保留低倍数显微图像基本信息的基础上，能够对高倍数显微图像细节进行有效的还原。

关键词: CycleGAN网络, 生成式对抗网络, 对抗学习, 卷积神经网络, 深度学习

Abstract: In terms of medical treatment,the diagnosis of many diseases relies on the observation of microscopic objects such as cells with a high magnification microscope.However,due to the high price and complex operation of high magnification microscope and there are some problems in the reconstruction of high magnification cell micro-images,such as the inconsistency of image style between high magnification micro-images and low magnification micro-images,the different resolution of cell images and the lacking of paired training data.To solve the above problems,a high magnification cell micro-images generative adversarial network is proposed.Based on the CycleGAN,a new residual dense block is added to the generator while the new activation function is introduced,and the Batch Normalization(BN) layers are removed.At the same time,in order to ensure the authenticity of the generated images,the detail perceptual loss is introduced to the training process of the generator.Experimental results show that the proposed method can effectively restore the detail of the high magnification micro-images while preserving the basic information of the low magnification micro-images.

Key words: CycleGAN network, Generative Adversarial Network(GAN), adversarial learning, Convolutional Neural Network(CNN), deep learning

中图分类号: 

• TP391