Stain Normalization Of Whole Slide Images Based On Generative Adversarial Networks

Globally,malignant tumors pose a serious threat to human health and life,and pathological diagnosis is the gold standard for clinical diagnosis.However,the shortage of pathologists,strong subjectivity,and low efficiency hinder pathological diagnosis.To improve the current situation,an effective solution is to digitize pathological sections through whole-slide scanning technology and then use computer-aided diagnosis to improve the accuracy and efficiency of pathological diagnosis.However,digital pathology images,also known as Whole Slide Images(WSIs),are affected by differences in tissue sample collection,staining methods,and scanning devices,resulting in significant differences in the staining styles of pathological images,even within the same medical center.These variations make automatic WSIs image analysis difficult.Therefore,there is an urgent need for pathological image stain normalization method to suppress staining variation between different pathological images,and improve the accuracy and generalization of computer-aided diagnosis.In recent years,Cycle-consistent Generative Adversarial Networks(CycleGAN)have been widely used for image style transfer due to their adversarial learning advantages.Therefore,this thesis explores the application of CycleGAN-based stain normalization for pathological images.The main research contents are as follows:(1)Aiming at the problem of staining variation within a medical center,this thesis proposes a stain normalization method named KStainGAN,based on a sparse clustering generative adversarial approach.Firstly,a sparse K-means stain feature clustering strategy is introduced to construct subsets of pathological images with consistent staining styles.Then,the CycleGAN architecture is utilized to achieve stain style transfer from diverse to consistent staining styles.Experimental results on a skin cancer dataset show that,compared with classical stain normalization methods,KStainGAN achieves better stain normalization results for pathological images,with a Structural SIMilarity(SSIM)of 0.96 and a Peak Signal-to-Noise Ratio(PSNR)of 32.62dB,versus 0.90 and 22.82dB,respectively.Additionally,when applied to assist in the diagnosis of skin cancer,the performance of the diagnostic model on KStainGAN normalized images is significantly improved,with an accuracy(ACC)of 90.70%and a macro F1 score of 94.90%,versus 76.74%and 84.30%,respectively,for unnormalized images.(2)Aiming at the problem of microstructure distortion and blocky artifacts caused by significant staining variation between different medical centers.This thesis proposes a stain normalization method named DRStainGAN,based on global stain disentangled representation generative adversarial approach.Firstly,a novel disentangled representation strategy is introduced in the CycleGAN generator,separating the microstructure and staining style features of pathological images,and guiding the network to learn the mapping relationship of staining style features to ensure microstructure consistency before and after normalization.Secondly,a global stain consistency loss is constructed to prevent blocky artifacts after normalization.Experimental results on MIDOG 21 and Camelyon 17 breast cancer public datasets show that DRStainGAN achieves better stain normalization results in maintaining microstructure(Contrast-Structure Similarity[CSS]:98.07 vs.80.55,SSIM:0.91 vs.0.70 and PSNR:21.42dB vs.17.13dB)and suppressing blocky artifacts(Global Stain Consistency[GSC]:3.33 vs.37.95)than state-of-the-art methods.Additionally,when applied to assist in the diagnosis of breast cancer,DRStainGAN normalized images significantly improve the F1 score of the diagnostic model by an average of 16.96%compared to unnormalized images.