Deep Learning Method For Urine Sediment Microscopic Image Segmentation And Classification

Urinary sediment examination has an important function in the diagnosis and differentiation of renal disease and urinary system disease,and is one of the routine inspection items in hospitals.Urinary sediment microscopic image analysis based on image processing is an important method of urinary sediment examination.It mainly includes segmentation and classification of urinary sediment particle images.Although,deep learning has made great progress in the field of image segmentation and classification in recent years,the urinary sediment particle images collected usually have problems such as defocusing,background noise,unclear outline of urine sediment particles and cell adhesion.The simple convolutional neural networks(CNN)segment the urinary sediment particle images will results in the segmentation edge not being smooth,with many noise,and the segmentation target is not complete.And because urinary sediment particle microscopic images have several different categories,small samples,and are affected by small intra-class differences and large within-class differences,which will cause the classification performance of CNN to be unstable.To this end,this paper first proposes a segmentation generative adversarial networks(SGAN)to segment urinary sediment particle microscopic images,and secondly proposes a classification method based on transfer learning and ensemble CNN for urinary sediment particle microscopic images,to provide reliable analysis result.It can be seen that the research on the urinary sediment particle images segmentation and classification based on deep learning has great social significance and academic value.The main research contents of this paper are as follows:(1)Investigate domestic and foreign scholars’ academic research on urinary sediment microscopic image segmentation and classification methods.Analyze the characteristics of urinary sediment microscopic images and urinary sediment particles,and summarize the technical difficulties of urinary sediment microscopic images segmentation and classification.In-depth study of the relevant theoretical basis of CNN and its application in the field of segmentation and classification.(2)Research on the segmentation method of urinary sediment particle microscopic images.There are problems such as unclear outline of urine sediment particles,cell adhesion,and defocusing in the urinary sediment microscopic images.Based on the idea of generative adversarial network(GAN),this paper designs a segmentation generative adversarial networks(SGAN)to segment urinary sediment particle microscopic images.The generative network is a 14 layer deep convolutional segmentation network constructed by residual structure and deconvolution structure,which is used to learn the distribution law of urinary sediment microscopic images,and generate segmented images.The discriminant network is a two-class network based on the residual structure,which is used to distinguish the difference between the generated segmented image and the label image,in order to distinguish the segmented image from the label image.The segmented image is closer to the label image by the adversarial training of the generative network and the discriminant network.At the same time,the two-class cross-entropy function and the perceptual loss are used as the loss function of the generative network to enhance the contour details of the segmented images.(3)Research on the classification method of urinary sediment particle microscopic images.There are problems such as many categories of urine sediment particles,similar particles,and lack of training samples.This paper proposes a classification method for urinary sediment particle microscopic images based on transfer learning and ensemble CNN.In view of the small dataset of urinary sediment microscopic images,which is not enough to train CNN,this paper uses the transfer learning idea to transfer the weights of the pre-trained CNNs on the ImageNet dataset to the urinary sediment microscopic images,and then fine-tune the CNN parameters such as learning rate to optimize its classification performance.In order to enhance the feature extraction ability of CNN,this paper first analyzes hierarchical nature of the features in CNNs by visualizing,then cascades the features of specific layers in CNNs.Because the redundancy of the features extracted by AlexNet,which affects its generalization ability,this paper proposes a developed AlexNet(De-AlexNet),which replaces the FC7 layer in AlexNet with two fully-connected layers(FCA1,FCA2),and reduces the feature size to 1024 dimensions.In the view of the advantages of a single CNN in feature extraction,this paper integrates the ResNet50,the GoogLeNet and the developed AlexNet to greatly improve the feature description ability for urinary sediment microscopic images.(4)Design a fully connected neural network for urinary sediment particles classification.Because the large feature dimension extracted by ensemble CNN and more redundant features,it may cause inaccurate classification.To this end,this paper constructs a Fully Neural Network(FNN)as a classifier to speed up classification while improving classification performance.(5)In order to verify the effectiveness of the segmentation and classification methods proposed in this paper,the urinary sediment microscopic image segmentation experiments were carried out under the framework of Tensorflow.The results show that the proposed segmentation method in this paper makes the segmentation edges smoother with better separation,and higher integrity.Under the Caffe framework,the urinary sediment microscopic image classification experiments were carried out.The results show that the classification method proposed in this paper greatly improves the classification accuracy of urinary sediment microscopic images.