Data Augmentation Of Gastric Tumor Cell Images Based On Generative Adversarial Networks

In recent years,as a common clinical disease,the incidence of tumors has gradually increased.Biopsy is an important basis for tumor diagnosis,which has shortcomings such as low efficiency and prone to human error.Therefore,an automatic diagnosis scheme is needed to assist doctors in tumor identification.Data-driven deep learning models are currently one of the best image recognition methods.Gastric tumor cell images,as a typical small sample data,require data augmentation before training.Traditional data augmentation methods such as rotation and cropping have limited effects,and it is difficult to meet the requirements of gastric tumor cell image data sets.The proposal and development of GANs have brought new directions for data augmentation.When using GAN to realize the augmentation of gastric tumor cell image data,there are problems of GAN training difficulty and slow training.Aiming at above problems,WGAN is used to generate gastric tumor cell images;in order to improve the effect of data augmentation,WGAN is improved to improve the quality of the generated image.Aiming at the problems of slow convergence and low recognition rate of Alex Net,Alex Net was improved and applied to the recognition of gastric tumor cell images.First,WGAN was used to augment the data of gastric tumor cell images.When GAN generates images of gastric tumor cells,there are problems of training difficulty and slow training.Using WGAN can effectively improve the above problems.WGAN uses the Wasserstein distance that reflects the distribution gap as the loss function and uses the gradient penalty to meet the Lipschitz continuity.Comparative experiments show that WGAN has a more stable training process and faster training speed,but data augmentation experiments show that the clarity of images generated by WGAN needs to be improved.Then,the WGAN is improved to improve the quality of the generated image.Because the long-term dependence of convolution is expensive,WGAN has limited ability to learn the positional relationship.Improve WGAN to increase the self-attention mechanism in the generator and discriminator to improve the network’s ability to learn the positional relationship;spectral normalization is applied for stabilizing the training process;Wasserstein divergence,a more theoretically complete loss function is used in order to remove the gradient penalty.The quantitative comparison with WGAN shows that the image generated by improved WGAN has higher definition;the experimental results show that improved WGAN has better data augmentation effect.At last,the improved Alex Net is used to recognize the augmented images of gastric tumor cells.Due to the small amount of Alex Net parameters and excellent performance in medical data,Alex Net is used to realize the identification of gastric tumor cell images.Aiming at the problems of slow convergence speed and poor generalization performance of Alex Net,it is improved by replacing the activation function and the normalization layer;and several experiments have been carried out to maximize the performance of the improved WGAN’s data augmentation.The experimental results show that the method of combining augmented data with improved Alex Net can well realize the recognition of gastric tumor cell images.