Glomerulus And Tubule Segmentation Model Based On Whole Kidney Puncture Image

Chronic kidney disease accounts for a large proportion of kidney diseases.Chronic kidney diseases can be roughly divided into primary and secondary kidney diseases,some of which can be divided into more kidney diseases respectively,such as Ig-A nephropathy,lupus nephritis and so on.However,the diagnosis and treatment of different chronic kidney diseases require pathologists to investigate different glomerular lesions in the kidney,such as sclerosis,crescent...;tubular lesions,such as tubular atrophy Therefore,achieving the segmentation of multiple types of glomeruli and multiple types of tubules on the complete renal puncture image will greatly improve the diagnostic efficiency of pathologists for chronic kidney disease.In order to achieve the segmentation of glomeruli and tubules on the complete renal puncture image,this paper designed different segmentation processes specifically.For the segmentation of multiple types of glomeruli,this paper proposed a strategy of "locating first,then segmenting".Firstly,the location of glomeruli was performed on the complete renal puncture image,and a reference box clustering algorithm based on sample quantity balancing was added to the glomeruli detection process,which provided more representative reference box sizes for the network.The experiment shows that the target detection network using the new reference frame clustering algorithm achieves 0.865,0.789 and 0.765 respectively for the glomerular detection on the complete image,and has a better performance compared with multiple networks.For the multi-class segmentation of glomeruli,this paper proposes a semantic segmentation network with dual task branching,which can separate the segmentation process from the classification process to ensure the integrity and unity of the segmentation results of glomeruli.At the same time,a loss function is creatively added to solve the problem of sample imbalance in the multi-class task.Experimental results show that the network and loss function proposed in this paper achieve the best results on multi-center data sets and have strong robustness and generalization ability.Average classification accuracy and F1-score showed the best results in three centers(Zhongda Hospital: 0.9,0.78;External verification set: 0.82,0.64;Gulou Hospital: 0.853,0.693),and the average split IOU was higher than other networks.For the segmentation of multiple renal tubules,this paper,according to the actual needs of pathologists,classifies the segmentation of renal tubules as the cortical part of the renal image,so as to construct the classification network of cortex,medulla and background,as well as the segmentation network of normal renal tubules and atrophic renal tubules.In the multi-class renal tubule segmentation network,a new multi-class segmentation loss function is proposed,which avoids the problem that the loss value is opposite to the actual result in special cases,and makes the network keep updating parameters to reflect the actual predicted result in the process of back propagation.Experiments show that the accuracy of the proposed tissue classification network reaches 0.983 on the three tissues.Compared with other loss functions,the loss function proposed in this paper achieved the best IOU for the two types of renal tubule segmentation,with an average IOU of 0.659.The segmentation system of glomeruli and renal tubules proposed in this paper based on total renal puncture images is based on the characteristics of each tissue structure and the actual needs of pathologists.It can greatly facilitate the analysis and judgment of pathologists for different structural lesions,and finally improve the diagnostic efficiency.