Research On Accurate Detection Of Urine Red Blood Cell From Multi Focus Videos Based On Deep Learning

Nephritis is a common major chronic disease.The patients often have urine occult blood and hematuria.The urine of patients contains normal and a variety of abnormal red blood cells.The proportions of different urine red blood cells(U-RBCs)are closely related to the type of nephritis.Therefore,detecting U-RBCs is the key to achieve non-invasive and reliable diagnosis of nephritis.Traditionally,completing the U-RBC detection can only depend on the experienced microscopists.However,the manual detection method has many interference factors,resulting in low accuracy,insufficient efficiency and limited scope of application.In recent years,relying on the single focus images obtained under the optical microscope,a series of U-RBC detection methods based on deep learning technology have been proposed.However,the images often contain numerous defocused U-RBCs,because the U-RBCs are distributed at different depths of the urine sample.The defocus phenomenon blurs and distorts the shapes of U-RBCs,which leads to the missed and false detection of existing methods.To overcome the limitations,a novel data type named multi focus video is proposed in the thesis.Relying on the multi focus videos,in order to effectively extract the features of U-RBCs and accurately detect them,the thesis designs three models,which are MI3D(Multi Instance inflated 3D CNN)model based on the multi model fusion strategy,RES-MIVi T(Rotation Edge Similarity Multi Instance Vision Transformer)model by combining the location method and MI3 D architecture,and MVF-RES-MIVi T(Multi focus Video Fusion Rotation Edge Similarity Multi Instance Vision Transformer)model by combining the rouleaux U-RBCs segmentation method and RES-MIVi T model.In addition,the thesis achieves the diagnosis of diabetic nephropathy based on the proposed HCD-KNN(Hierarchical clustering Cluster Deviation K-Nearest Neighbor)classification model and the proportion of detected U-RBCs.The main innovations are introduced as follows:(1)The proposed MI3 D model can accurately learn the complex and confusing features of U-RBCs from the multi focus videos.The multi focus videos of U-RBCs contain two effective features: continuous time features and typical space features.The continuous time features demonstrate the similar deformation regularity of same class U-RBCs,and the typical space features is the easily recognized U-RBC shapes.The MI3 D model integrates the Inflated 3D convolutional neural network(I3D)model to learn the deformation regularity of U-RBCs,and the Multi Instance Learning(MIL)model to learn the typical shapes of U-RBCs.The fusion mechanism represented by significance pretraining ensures that the multiple models in MI3 D can be synergistically trained.The experiments prove that the accuracy of MI3 D model achieves 0.944,which is 8.8% to 20.0% higher than the existing methods and lays a solid foundation for the subsequent detection methods.(2)The proposed RES-MIVi T model achieves the accurate detection of U-RBCs from multi focus videos.This model effectively overcomes the missed and false detection problem caused by the defocused U-RBCs.The RES-MIVi T model contains two stages:location and classification.In the location stage,the rotation edge similarity(RES)method can determine the accurate edges of U-RBCs in any frame and segment the multi focus video into a series of partial videos containing only one U-RBC.In the classification stage,MIVi T is responsible for classifying the partial videos in the location stage and outputting the accurate detection results.The MIVi T is a balanced feature extraction model obtained by replacing the I3 D network in MI3 D with the video frame encoding vision transformer network.In the experiments,the mean average precision of RES-MIVi T achieves 0.939.(3)The proposed MVF-RES-MIVi T model accomplishes the accurate detection of rouleaux U-RBCs from multi focus videos.This model overcomes the limitation of RES-MIVi T which cannot effectively detect the adhesive and encysted U-RBCs.The construction strategy of MVF-RES-MIVi T model is to add a fusion stage on the basis of the RES-MIVi T model.In the fusion stage,the three priori indicators are used to extract the effective shapes of U-RBCs and fuse them into images.Then,in fused images,the rouleaux U-RBCs are separated by the YOLOv4 model.The separated U-RBCs are located by the RES method and classified by the MIVi T model.In the experiments,the mean average precision of MVF-RES-MIVi T achieves 0.954,which is 25.4% to 29.3% higher than mainstream methods.(4)The proposed HCD-KNN model accurately classifies the proportion of normal and heterocyst U-RBCs detected by the MVF-RES-MIVi T model,and accomplishes the reliable diagnosis of diabetic nephropathy.The existing methods generally use the inaccurate artificial threshold to judge the correlation between the proportion of U-RBCs and diabetic nephropathy.By adding the hierarchical clustering simplification and cluster deviation mechanism for K-nearest neighbor method,the HCD-KNN model is constructed.The HCD-KNN model effectively analyzes the proportion regularity of U-RBCs and obtains0.807 diagnosis accuracy.In conclusion,the models in this thesis can assist the microscopists completing the accurate U-RBC detection and reliable diabetic nephropathy diagnosis.