| ObjectiveIn recent years,with the aging of the population,environmental pollution,the incidence of membrane nephropathy is getting higher and higher,the current diagnosis of membrane nephropathy mainly depends on kidney biopsy,the final clinical diagnosis still needs pathological guidance treatment.Subsequently,the increase in the amount of pathological reading,increased the burden of pathologists,but also increased the investment of human and material resources of medical institutions.Artificial intelligence assistive technology gradually appears in people’s vision,whether it can use the deep learning method in artificial intelligence technology,establish a computer-aided diagnostic system,realize the automation and rapid diagnosis of membrane nephropathy,reduce the workload of clinical pathologists,improve the efficiency of clinical diagnosis and treatment,rationalizing the allocation of medical resources,is the main content of this research.MethodUsing retrospective data analysis method,this study selected 216 patients who met the criteria for kidney biopsy from March 2015 to June 2020 at the Renal Hemodrosis Center of Shaanxi Provincial People’s Hospital,and collected 5,964 kidney glomerular balls in the kidney pathology picture.Among them,the positive group was 189 cases of membrane nephropathy(Phase I 96,Phase I 66,Phase III 26,Phase IV 1),and the negative control group was 27 cases of mild lesions of the renal glomerular.Of these,139(64.4%)were men and 77(35.6%)were women,with age ranges of 15-88 years,and all patients included in the study collected four dyed pathological tissue slices from Masson,HE,PAS,and PASM,respectively A total of5,964 standard-compliant renal trophic balls were collected under × 200 times the optical mirror,of which 3,269 were in the membrane nephropathy group,735(12%)in Phase I and 1190(20%)in Phase II.Phase III 1344(23%);Two sets of different staged kidney glomerular pictures with Anaconda 3.0 mark software labeling,and with different coding marker stage,the already phased labeled kidney glomerular pictures into the computer for deep learning processing,late without label input computer,using the "black box" principle,to see whether the machine output pathological classification and diagnosis in line with the pathologist diagnosis results,compare the accuracy of the results.ResultsIn this study,the positive group and the negative group marked a total of 3152 pictures of kidney glomerulars,including the number of kidney glomerulars 5964740),3269 renal glomerulars(Phase I 735,Phase II 1190,Phase III 1344);The average correct rate(m AP)of different stage kidney glomerulars was 0.88 at the beginning of the deep learning model,of which Phase I 0.80,Phase II 0.81,Phase III0.96,and the mild lesions group of renal glomerulars were 0.94.Conclusion1.Compared with the accuracy of manual markup,the computer-aided diagnostic system based on deep learning algorithm has an average accuracy rate of0.88 for different stages of diagnosis of membrane nephropathy,and has a high diagnostic accuracy.2.Through this study,the diagnostic data set of membrane nephropathy was initially established,the distribution of different phased data sets was uneven,the number was too small,and the data needed to be updated and perfected later in life,so as to further establish a complete database.3.The system is also suitable for other clinically related diseases that rely on pathological diagnosis and has a wide range of applications. |
PDF Full Text Request