Artificial Intelligence-assistid Diagnosis For Electron Microscope Image Of Renal Biopsy

Background and PurposeGlomerulonephritis is a common kidney disease in China and the main cause of chronic renal failure,and the most common pathological type is thylakoid proliferative glomerulonephritis.The onset of glomerulonephritis is relatively insidious,characterized by high incidence and low awareness,and the long course of the disease imposes economic and mental burdens on patients,so early and accurate diagnosis and treatment are essential for the prevention and treatment of glomerulonephritis.Renal puncture biopsy is the main diagnostic modality to clarify renal disease.In clinical work,pathologists need to combine light microscopy,immunofluorescence and Transmission electron microscope(TEM)images of kidney puncture biopsy tissues and clinical information for comprehensive analysis before giving the final diagnosis.In order to observe the ultrastructure of the glomerulus,pathologists need to measure the thickness of the glomerular basement membrane,calculate the number of foot processes or the degree of fusion,and clarify the deposition site of electron dense deposits,which is tedious and energy-consuming.In recent years,the rapid development of Artificial Intelligence(AI)combined with medical imaging has provided us with new ways to solve these problems.In this study,we propose to use an innovative algorithm to automatically segment and identify glomerular basement membrane,foot processes and electron dense deposits in TEM to assist pathologists in diagnosis and improve diagnostic efficiency.MethodsIn this project,693 TEM images of 349 patients with glomerulonephritis were collected and screened based on the inclusion criteria of TEM images with clear glomerular structure and magnification of 4000×.The pathological types included IgA nephropathy(74 cases),membranous glomerulonephritis(57 cases),thin glomerular basement membrane disease(26 cases),minimal change disease(58 cases),mesangial proliferative glomerulonephritis(53 cases),lupus nephritis(48 cases),and diabetic nephropathy(33 cases).The 3 components of glomerular basement membrane,foot processes and electron dense deposits in TEM images were labeled by pathologists,and the basement membrane,foot processes and electron dense deposits images were divided into training and test sets according to the ratios of 8:2,6:1 and 8:2.(1)Automatic segmentation and identification of basement membrane in glomerular TEM images using PraNet model.(2)Automatic segmentation and identification of foot processes in glomerular TEM images using the optimized Unet3+model.(3)To construct a Co-Net model for automatic segmentation and identification of electron dense deposits in glomerular TEM images.Results(1)The average Dice coefficient of the PraNet model for automatic segmentation and identification of the basement membrane was 0.782 on the test set,and the model predicted the smooth and regular basement membrane well.The segmentation accuracy of the model needs to be improved in case of severe morphological changes in the basement membrane,such as the presence of electron dense deposits embedding or substantial distortion of the basement membrane.(2)Using the optimized Unet3+model for automatic segmentation and identification of foot processess,the average Dice coefficient on the test set was 0.650.The model predicted the location of foot processess well,but the prediction of foot processes morphology was not yet satisfactory,and the structure of the model still needs to be optimized to improve the accuracy.(3)The average Dice coefficient of the Co-Net model for automatic segmentation and identification of the electron dense deposits was 0.797 on the test set.The model predicted the location of electron dense deposits well,but could not predict the edges and irregular regions of electron dense deposits completely.Conclusion1.In this project,we successfully established PraNet model,optimized Unet3+model and Co-Net model,and the accuracy of identification and prediction of basement membrane,foot processes and electron dense deposits in glomerular TEM images reached over 65%,which laid a good foundation for our later work.2.Subsequently,we will optimize the structure of our model,improve the performance of the model,increase the data set,further improve the accuracy of basement membrane,foot processes and electron dense deposits prediction,provide valuable lesion information,assist pathologists in diagnosis,improve diagnostic efficiency,and reduce the workload of pathologists.