Study On Image Transformation Of Tissue Section Staining Based On CycleGAN

Objective: Hematoxylin-eosin staining and IHC staining are commonly used in cell staining.However,IHC staining has the disadvantages of complex operation and unstable staining results,while hematoxylin-eosin staining is based on the simple operation of acidbase neutralization.Therefore,we used image style transfer to translate the results of HE staining into the IHC results of specific proteins.The objective of our experiment is to study a generative antagonistic network for image style transfer of hematoxylin-eosin stained cytoplasmic sections and IHC stained images,improve the clarity of the generated images and reduce the noise in the images.Methods: This research is mainly based on cyclic adversarial network.Continuous sections of mouse kidneys were selected for hematoxylin-eosin staining and IHC staining of different proteins.IHC images of three proteins were selected for testing,and the size of the images was reduced to 256×256 for testing.We design this model based on cyclic generation adversarial network.A multi-scale convolution channel is designed in the encoder part of the generator,and attention gating is added in the feature fusion.Through this method,the details of the generated images are enhanced and the clarity of the generated images is improved.We used evaluation indicators FID,PSNR,SSIM and VIF to verify the quality of the images we generated.Results: A total of 4 staining images of cell sections were used in this study,including one HE staining image and three IHC staining images of proteins.After each picture is cut to256×256,each set of images is divided into training set and test set according to 8:2.And compared with the mainstream image style conversion algorithm,finally calculate the evaluation index.For the metric PSNR,the scores for the three protein conversion results are as follows 28.085,28.163,28.296,respectively.For the evaluation index SSIM,the conversion results of the three proteins were 0.81,0.674 and 0.527,respectively.Qualitative and quantitative evaluation and analysis of experimental results,as well as subjective observation of generated images.we can conclude that our network has well completed the research task of transforming the cell staining sections from HE staining images to IHC staining.Conclusion: In this study,an image style migration algorithm by generating adversarial network was developed based on the stained cell slice images to transform the cell slice images stained by HE into those stained by IHC.By designing a GAN generator,we can make our model have better performance.The generator in this experiment has the convolution of two channels in the feature extraction part,and the attention gating mechanism is added in the feature fusion part,so that the details of the generated pictures are richer and the clarity of the pictures is improved.