Preliminary Study On The Recognition Of Glomerulus And The Classification Of Spikes In Pathological Images Of Kidney Tissue Based On Deep Learning

Membranous nephropathy(Membranous nephropathy,MN)is one of the common pathological types that cause nephrotic syndrome(Nephrotic syndrome,NS).In recent years,the incidence of membranous nephropathy has gradually increased,and some patients have progressed to end-stage renal disease or died of related complications within 5 to 10 years.The diagnosis of membranous nephropathy mainly depends on renal biopsy.However,there is an imbalance in the distribution of medical resources in China.High intensity film reading by pathologists not only consumes a lot of energy,but also temporarily reduces the working ability due to work fatigue,leading to a decrease in diagnostic efficiency.The current digital pathology technology appears in the medical field,especially the deep learning model convolutional neural network in image recognition and processing,and the pathological diagnosis system based on computer image processing has brought a broader development space.Therefore,this research through the entire glomerular pathological slides to all the information and complete a quick scan,kidney pathological image by digital neural network as the core to develop automatically on glomerulus in renal tissue pathological image recognition software systems,able to quickly identify glomerular and accurate count.Subsequently,the microscopic structure of the pathological nail process inside the glomerular lesion was analyzed from the pathological sections of renal tissue,so as to reduce the misdiagnosis or missed diagnosis caused by the subjective factors of physicians(such as lack of clinical experience or work fatigue).To help the professional kidney pathologist to improve the diagnostic efficiency of the purpose.Part Ⅰ Recognition and localization of glomeruli in pathological images of renal tissue based on Cascade R-CNN network Objective:Artificial Intelligence based on deep learning has been rapidly developed in the field of pathological detection,and glomerulus is the main pathological lesion site of membranous nephropathy.This study is based on the Cascade R-CNN network to develop an artificial intelligence(AI)system that can automatically identify glomeruli in pathological slice images of renal tissue,helping pathologists to improve the efficiency of calculating the number of glomerulus and identifying glomerulus.Methods:Collect 1,250 renal biopsy pathological slices of patients with membranous nephropathy in Shanxi Provincial People’s Hospital affiliated to Shanxi Medical University and the Second Hospital of Shanxi Medical University.After scanning and saving,the pathological images that meet the requirements are screened and converted into a format recognized by the machine.After the picture is preprocessed,it is imported into the artificial intelligence recognition system developed by the Cascade R-CNN network,and the recording software automatically recognizes the glomerular precision and recall rate and F1.The pathological images were read by three pathologists with more than 3 years of work experience,and the precision and average time for the physicians to identify the glomeruli were collected.Analysis of variance was used to analyze whether there were statistical differences in the recognition time between pathologists;the chi-square test was used to analyze whether there were statistical differences in the precision of identifying glomerulus among pathologists.The independent sample T test was used to analyze whether there is a significant difference between the Cascade R-CNN network model and the pathologist’s accuracy in identifying Glomerulus.Results:1.Data set analysisA total of 1150 pathological images were selected to meet the requirements.5226 glomerular area images of 3000×3000 pixels were obtained by overlapping and cutting920 pathological images(80%)as a training set,and the bounding frame of 40481 glomeruli labeled by pathologists was finally determined.230 pathological images(20%)were overlaps and sected to obtain a test set of 1244 glomerular region images of3000×3000 pixels.2.Quality testingAfter analyzing the precision and recall graphs,it is clear that the model trained by the Cascade R-CNN network is very expressive in the test set to detect the features obtained by automatic learning of the glomerulus.3.Comparison of Cascade R-CNN network and Faster R-CNN networkIn the early stage,the accuracy of identifying glomeruli based on Faster R-CNN uses Mean Average Precision(m AP)to measure the performance of each model.The m AP value is 94.37%,and the image processing time of the entire slide is about 1 s.the experiment results show that the Cascade R-CNN deep learning model of network training to complete processing the whole piece of glass image of time is about 1 s.Cascade R-CNN is a target detection algorithm combining several different IOU thresholds on the basis of Faster R-CNN.In terms of detection structure,feature extraction takes the most time.According to the results of this experiment,although Cascade R-CNN adds more parameters,it has no obvious influence on the detection speed.According to the test results of the model on the test set,the precision and recall rate of our algorithm model are 94.50%,98.07%,and F1 is 96.25% respectively.4.Comparison of Cascade R-CNN model and pathologist’s diagnosis resultsThe time for the three pathologists to recognize each cut pathological image was3.57 ± 0.05 s,4.52 ± 0.07 s and 3.98 ± 0.02 s,respectively;there was no significant difference in the time for the pathologists to recognize each cut pathological image by ANOVA.The precision of glomerular identification among the three pathologists was88.08%,89.69%,and 89.98%.The chi-square test showed no statistical significance in the accuracy of glomerular identification among pathologists.The recognition time of each glomerular region image by Cascade R-CNN network was 0.20±0.02 s.The precision of the Cascade R-CNN network for glomerular identification was 94.50%,and the average precision of pathologists for glomerular identification was 89.25%.The independent sample T test was used,and the difference between the two was statistically significant(t=-5.61 P=0.01).Conclusion:The Cascade R-CNN network can quickly identify glomeruli through high-resolution WSI.Based on the Cascade R-CNN network to develop an artificial intelligence(AI)system,it can help pathologists improve the efficiency of calculating the number of glomeruli and identifying glomeruli.Part Ⅱ Classification of glomerular spikes in pathological images of membranous nephropathy based on Goog Le Net network Objective:To evaluate the classification of glomerular spikes in membranous nephropathy by Goog Le Net Inception V1 network,so as to help pathologists find the tiny structure of spikes in the pathological detection of membranous nephropathy,and assist pathologists in assisting the diagnosis of membranous nephropathy.Materials and Methods:From the 1150 PASM stained pathological images in the first part,127 pathological images were extracted pathologically diagnosed as stage II and stage III of membranous nephropathy.The glomeruli were identified and segmented by Cascade R-CNN network.Five pathologists classified the glomeruli after cutting.The glomeruli with spikes were put into a folder,named as spikes,with a total of 517;the glomeruli without spikes were put into a folder,named as no spikes,with a total of 498.No image-based labeling was performed in the process.The training set contains 710 images,the validation set contains 204 images,and the test set contains 101 images.We use the calculation results of the test set to evaluate the performance of the model.Import into Goog Le Net Inception V1 network for training.The accuracy,Precision,recall rate,and F1 value are used to evaluate the performance of the Goog Le Net network in the classification and detection of glomerular spikes.1.Data set analysisA total of 127 PASM stained pathological images of patients with clearly diagnosed membranous nephropathy stage II and stage III were collected.The glomeruli were identified and segmented through the Cascade R-CNN network to obtain 1015 glomeruli.There were 5-21 glomeruli on each PASM stained pathological image,and 1-16 glomeruli with spikes on each PASM stained pathological image.2.Quality InspectionThe performance evaluation of the Goog Le Net Inception V1 model is based on whether there is a glomerulus with spikes that is correctly detected and a binary classification model.According to the test set results,the recall rate is 97.96%,the Precision is 85.71%,the accuracy is 91.09%,and the F1 is 91.43%.A total of 101 glomerular images in the test set were detected,with a total time of 298.32 s.According to the loss and validation accuracy curves of the model during training,the loss decreases and the accuracy improves,which verifies the convergence of the model and the effectiveness of model training.Conclusion:Goog Le Net Inception V1 model can well classify the spikes and obtain a high recall,but the precision and accuracy need to be further improved,so as to provide an important reference for pathologists in the diagnosis of membranous nephropathy.