Research On Unsupervised Generative Adversarial Methods For Re-staining Structural Pathological Images

Histopathological examination is the gold standard for the clinical diagnosis of cancer.Pathologists first extract the biopsy tissue of suspected human lesions and then use staining and other methods to prepare tissue samples.Finally,high-resolution histopathological whole slide images(WSIs)can be obtained through high-precision instrument scanning and multi-field seamless splicing.The traditional tissue staining process is irreversible,so it is challenging to construct paired pathological image datasets.Secondly,pathologists need to monitor the entire staining process under a microscope,and the number of pathologists in China is scarce,and biopsy tissue is expensive,so it is time-consuming and labor-intensive to prepare different types of histopathological WSIs.Based on the background,this paper studies unsupervised generative adversarial methods for re-staining structural pathological images,that is,using the existing WSIs to obtain other structural staining WSIs required on the premise that the medical interpretation of the images remain unchanged to achieve the purpose of reversible staining process and saving staining cost.Compared with natural scene images,pathological WSIs have the characteristics of high-resolution,complex and diverse cells or tissues,and difficulty in obtaining paired data.Therefore,current re-staining work generally have problems such as tiling artifacts and staining diffusion.In order to solve staining diffusion,this paper studied the restaining of pathological images from two dimensions of preserving content features and promoting local regional style transfer,and achieved the following research results:Firstly,a virtual re-staining network is proposed based on contrastive learning and high-level losses guidance.In order to ensure that the content features such as the size and shape of cells or tissues remain unchanged,we model based on the dual multi-scale contrastive learning framework and combine cyclic structure consistency loss,perceptual embedding consistency loss,and multi-scale contrastive learning loss to preserve the consistency of the original structure,the high-level semantic features and the multi-scale latent features extracted by the encoder.Experiments show that these modules can effectively improve the performance of pathological image re-staining and make the proposed model outperform all baseline models.Secondly,a multi-scale style disentangled virtual re-staining network is proposed for target domain images.The current circular frameworks can only transfer the global style well but have weak constraints on the style transfer of local regions.Therefore,this paper adds a multi-scale style disentangled branch to the unilateral model of Study 1.Specifically,the disentangled style features are extracted through the style encoder and the feature mapping network using the target domain image as the reference.Then through different affine layer mapping and adaptive instance normalization operations,the style guidance information of different scales can be provided for the re-staining process,which can promote the network to transfer the global style while ensuring that the style features and content features of local regions are well integrated.Experiments show that the proposed model can convert pathological images closer to the distribution of target domain data.