Research On Virtual Staining Method Of Carotid Artery Sections Based On Conditional Generative Adversarial Networks

Cardiovascular disease has become an important disease threatening human health.Stent implantation in the process of treatment will cause damage to the inner wall of blood vessels and lead to intimal hyperplasia,resulting in vascular restenosis and other problems.At present,the physiopathological researches of vascular intimal hyperplasia mainly rely on the histological analysis of pathological sections,which is also the gold standard for the diagnosis of cardiovascular diseases.However,the staining of pathological sections is a time-consuming and tedious process.The chemical reagents used can also cause irreversible damage to the tissue or even lead to misdiagnosis.The inconsistency in staining depth of pathological slices reduces the reuse rate of sections.In this paper,a virtual staining method of rat carotid artery unstained pathological sections has been researched under the above background,the main work are as follows:(1)An algorithm for virtual staining prediction of unstained pathological slices of blood vessels based on conditional generative adversarial networks(c GAN)was proposed.The microscopic image of unstained pathological section was input into the trained network,a staining pathological section which was similar to the standard staining could be output.Using machine learning methods to directly obtain clinical staining equivalents from unstained sections of blood vessels,bypassing labor-intensive and expensive tissue staining procedures,has important research significance.The chemical staining methods used in this paper were hematoxylin-eosin staining(H&E),picrosirius red staining(PSR)and orcein staining(Orcein).(2)Pre-processing of rat carotid artery tissue samples provided by the hospital,including taking blood vessel slices under a microscope to obtain RGB images,was executed before training.Based on the image registration algorithm,an unstained slice and standard stained section were registered,the data volume was cropped to expand training data and testing set for a total of 5300 cases.(3)According to the requirements of clinical diagnosis,a blind evaluation index of doctors conforming to the characteristics of this pathological data was formulated.The pathologists evaluated virtual staining results generated by the network and standard staining sections on the basis of score table.The evaluation results of pathologists proved that there was no significant difference between virtual staining and standard staining.At the same time,structural similarity(SSIM)and peak signal-to-noise ratio(PSNR)were used to quantify the network learning ability.SSIM reached 0.915 and PSNR was 26.43.(4)Different network structures,including WGAN and DCGAN,were compared on this pathological dataset.The combination of c GAN and U-net proposed in this paper had the strongest performance in details finally.WGAN learning overall fuzzy,SSIM was 0.783.DCGAN was blurred in the contour,SSIM was 0.682.(5)In order to improve the learning efficiency of pathological sections,a generator structure based on Star GAN network was proposed to realize the learning algorithm of multiple staining intervals.A pathological section to be learned was input into Star GAN,the unstained pathological section,H&E staining,PSR staining and orcein staining could be output simultaneously.However,compared with the results of c GAN,the structural details of learned pathological sections need to be optimized,SSIM was 0.6.The proposal of virtual staining improves the traditional pathological staining process of vascular tissue and saves a lot of human resources and time cost.It can be used as a blueprint for virtual staining of tissue images and other label-free imaging modes to improve the reuse rate of pathological sections,which has important potential clinical application value.