Deep-Learning Based Tissue Segmentation And Stain-Style Transfer Methods On Pathology Image

Oropharyngeal cancer is a malignant tumor originating from the epithelial tissue of the oropharynx,and its incidence and mortality rates have been on the rise globally in recent years,making it a significant public health issue.Precise segmentation of the tumor epithelial tissue area in patient pathological slices can help formulate individualized treatment and follow-up plans and holds great clinical significance.However,pathological features of oral-pharyngeal cancer are heterogeneous,and tumor epithelial tissue can exhibit similar image features as other tissues such as tumor stroma in HE-stained images.Additionally,pathological images generated by different hospitals and equipment have significant variations,increasing the difficulty of epithelial tissue segmentation tasks in pathological images.Compared to HE-stained images,IHC-stained images can significantly improve the efficiency and accuracy of epithelial tissue segmentation,but obtaining IHC-stained images has high costs and is time-consuming.Therefore,there is an urgent need for a method that can convert existing conventional stained images into corresponding IHC-stained images.To address the above problems,this study conducted the following work:(1)For the pathological image epithelial tissue segmentation task,this study proposed a segmentation model named Epi UNet.This model has strong feature extraction ability by introducing a multi-level nested sub U-Net structure and deep supervision strategy to extract deeper feature information and eliminate irrelevant features,achieving a balance between segmentation accuracy and computational complexity.In addition,this study proposed an innovative loss function to capture spatial similarity and boundary features between patches,and the effectiveness of Epi UNet was ultimately demonstrated.(2)For the pathological image staining conversion task,this study proposed a staining style transfer model named CS-Net based on cycle GAN.The generator and loss function of cycle GAN were improved,and an attention module named CS-Gate was introduced in the skip connection of each layer to extract multidimensional feature information of pathological images and weaken unnecessary features.The experimental results showed that CS-Net can synthesize higher quality IHC-stained images and outperforms existing state-of-the-art methods on multiple evaluation metrics.(3)Combining the above two works,this study proposed a two-step epithelial tissue segmentation framework based on staining style transfer.In the first step,the HE-stained image to be segmented is converted into a corresponding pseudo IHC-stained image using the previously proposed staining style transfer network.In the second step,the binary mask of the epithelial tissue is obtained by applying the image processing algorithm proposed in this study.The final comparative experiments verified the effectiveness of this method,and the performance of the obtained mask was even better than the current best segmentation model,providing a new solution for the epithelial tissue segmentation task in pathological images.