{"title":"Single Generative Networks for Stain Normalization and Quality Enhancement of Histological Images in Digital Pathology","authors":"Xintian Mao, Jiansheng Wang, X. Tao, Yan Wang, Qingli Li, Xiufeng Zhou, Yonghe Zhang","doi":"10.1109/CISP-BMEI53629.2021.9624221","DOIUrl":null,"url":null,"abstract":"Stain normalization of histopathology images is a promising technique commonly used in computer-aided diagnosis. This process eliminates the effects of staining intensity and color difference (batch effects) from various pathologic imaging systems. In this paper, we are focusing on stain normalization and visual quality enhancement. Although state-of-the-art methods, such as CycleGAN, perform well in image style transfer, they have been limiting by raw imaging quality. This paper propose a novel framework, single generative networks (SGNet), to train the staining model. We yield data pre-augmentation instantiated by clarity-brightness-saturation (CBS) adjustment, and introduce max pooling between the input and the intermediate features and positional normalization (PONO) to optimize network structure. The proposed approach is evaluated by using the placental pathological samples with villi, trophoblast cells and vascular area. Feature fusion results on placental sample demonstrate the proposed model outperforms existing methods, ESPCN, CycleGAN and SegCN-Net. Ablation studies also show the necessity of additional components. We test this network on low-quality images from different imaging systems. Experimental results preserve detailed structural information of tissues and show desirable performances on generalization ability of histological image, which increases the segmentation accuracy for digital pathology diagnosis. These findings have the potential for the establishment of histological staining criterion, massive pathological images with batch effects can be normalized with the aid of authoritative staining benchmark.","PeriodicalId":131256,"journal":{"name":"2021 14th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI)","volume":"11 1","pages":"0"},"PeriodicalIF":0.0000,"publicationDate":"2021-10-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"2","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"2021 14th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI)","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/10.1109/CISP-BMEI53629.2021.9624221","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"","JCRName":"","Score":null,"Total":0}
引用次数: 2
Abstract
Stain normalization of histopathology images is a promising technique commonly used in computer-aided diagnosis. This process eliminates the effects of staining intensity and color difference (batch effects) from various pathologic imaging systems. In this paper, we are focusing on stain normalization and visual quality enhancement. Although state-of-the-art methods, such as CycleGAN, perform well in image style transfer, they have been limiting by raw imaging quality. This paper propose a novel framework, single generative networks (SGNet), to train the staining model. We yield data pre-augmentation instantiated by clarity-brightness-saturation (CBS) adjustment, and introduce max pooling between the input and the intermediate features and positional normalization (PONO) to optimize network structure. The proposed approach is evaluated by using the placental pathological samples with villi, trophoblast cells and vascular area. Feature fusion results on placental sample demonstrate the proposed model outperforms existing methods, ESPCN, CycleGAN and SegCN-Net. Ablation studies also show the necessity of additional components. We test this network on low-quality images from different imaging systems. Experimental results preserve detailed structural information of tissues and show desirable performances on generalization ability of histological image, which increases the segmentation accuracy for digital pathology diagnosis. These findings have the potential for the establishment of histological staining criterion, massive pathological images with batch effects can be normalized with the aid of authoritative staining benchmark.