Segmentation And Recognition Of Chronic Kidney Disease Electron Microscope Image Based On Deep Learning

Background and purposeChronic kidney disease(CKD)is one of the diseases that seriously affect human health,and because there are currently no effective drugs and other means for radical treatment,it causes a huge economic burden and mental stress to families and society,so it should be detected,diagnosed and treated early.At present,renal biopsy is the main method for the diagnosis of chronic renal disease.By means of Transmission electron microscopy(TEM),Pathologist can observe not only the ultrastructure of the kidney tissue,but also the presence or absence of electron dense bodies and location of their distribution,combined with light microscopy and immunofluorescence,so as to make further pathological diagnosis.In the process of pathological diagnosis,the pathologist needs to identify the glomerular basement membrane and measure its thickness,calculate the number and average width of foot processes,and identify the electron dense deposits and location of its distribution.However,in TEM images,the texture of the glomerular basement membrane,foot processes and electron dense deposits are complex,their shapes are diverse,and their and surrounding structures are not easily distinguished,if only through naked eye to TEM image analysis is not only time-consuming but also inaccurate.Therefore,we propose a deep learning method for automatic segmentation and recognition of glomerular basement membrane,foot processes and electron dense deposits,which is helpful to improve the diagnostic efficiency of renal pathologists for chronic kidney disease.MethodsAccording to the inclusion criteria,the TEM images of 326 patients were collected,and the glomerular basement membrane,foot processes and electron dense deposits were labeled.According to the ratio of 9:1,9:1 and 8:2,the images of basement membrane,foot processes and electron dense deposits were divided into training set and test set.(1)In this study,we use PraNet algorithm to segment and recognize glomerular basement membrane automatically;(2)For automatic segmentation and recognition of foot processes,we use Unet3+model;(3)We use PraNet algorithm to segment and recognize electron dense deposits in TEM images automatically;Finally,the automatic segmentation and recognition of glomerular basement membrane,foot process and electron dense deposits were compared with manual labeling,and the accuracy was quantitatively evaluated from the visual effect and Dice coefficient.Results1.In the automatic segmentation and recognition of glomerular basement membrane based on PraNet algorithm,the average Dice coefficient on the test set is 0.770;2.In the automatic segmentation and recognition of foot processes based on Unet3+model,the average Dice coefficient on the test set is 0.792;3.In the automatic segmentation and recognition of electron dense deposits in TEM images based on PraNet algorithm,the average Dice coefficient on the test set is 0.725.4.In addition,each TEM image can be segmented and recognized in less than 1 second by various segmentation and recognition models.ConclusionsThe automatic segmentation and recognition method proposed in this paper has achieved good results in the automatic segmentation and recognition of glomerular basement membrane,foot processes and electron dense deposits.The accurate segmentation and recognition lays a foundation for the later measurement of glomerular basement membrane thickness,the calculation of foot processes number and average width,and the judgment of electronic density distribution,It also provides valuable information for renal pathologists in the diagnosis of chronic kidney disease.