A Method Based On Multiple Instance Learning For Glomerular Classification And Lesion Visualization

Membranous nephropathy（MN）is a common type of nephrotic syndrome in adults,with an incidence of 23.4%.Renal puncture biopsy is the gold standard for the diagnosis of membranous nephropathy.The traditional pathological diagnosis method requires doctors to judge the degree of immune complex deposition on the glomerular basement membrane under a microscope,and such a huge workload is very likely to lead to misdiagnosis and omission.However,there are still big problems in the practical application of computer algorithms.Most of the lesions are very small in the glomerulus.If the lesion is segmented at the pixel level,it often relies on a large number of pixel-level labels.Many tiny lesions make pixel-level labeling extremely difficult and pose a great challenge for model training.In order to solve these problems and achieve the purpose of auxiliary diagnosis of membranous nephropathy.We establish a deep residual network model based on a multi-instance learning framework to achieve glomerular classification in renal pathological images.Firstly,U-Net is used to extract the region of the glomeruli to ensure that the features learned by the succeeding algorithm are focused inside the glomeruli itself.Secondly,we use multiple instance learning（MIL）to train an instance-level classifier combined with multi-scale annotation method to enhance the learning ability of the network via adding a box-level labeled reinforced dataset,so as to obtain an instance-level feature representation with fine-grained feature information.In the test phase,the predicted probability of each block in the glomerular image is summarized,so as to give the classification result of the glomerulus,and obtain the segmentation and positioning result of the lesion in the glomerulus.At the same time,the quantified index of the nail process of typical membranous nephropathy was constructed on the heat map,and the ratio of the area of the nail process to the area of the glomerulus was calculated,and the relationship between this index and the stage of membranous nephropathy was found to achieve the quantitative analysis.In this paper,a total of 1267 glomeruli on 222 full-field pathological slices were collected,and the proposed model was trained and tested using these data.In the glomerular segmentation stage,the model can effectively segment the glomerular edges with an average dice coefficient of 0.912.In the glomerular classification stage,the multi-scale annotation-based deep multiple instance learning model constructed in this paper can quickly converge,and it has better classification performance compared with other models.The ₁F-measure of glomerular classification reaches 0.9580.In the visualization stage of the lesion,this model can accurately give the location of the spike in the glomerulus.The proportion of the area of the glomerulus is used as a measure of the number of spikes,realizing the quantitative analysis of spikes in the glomerulus.The model proposed in this paper can automatically classify glomeruli and visualize lesions in kidney pathological slices,and complete quantitative analysis tasks,providing doctors with a reference basis for pathological diagnosis,greatly reducing the workload of doctors,and automating the computer.The proposed model can provide a good foundation for assisting the clinical doctors to diagnose the glomerular membranous nephropathy.