Classification And Detection Of Smear Cell Images Based On Convolutional Neural Network

Cytology smear plays a key role in the diagnosis,treatment,and prognosis of clinical diseases,and observing the morphology,number,and other characteristics of smear cells helps medical experts to assess the patient’s condition,which is intuitive,quick,and accurate.However,the cytological smear contains different cell morphologies and complex cell-to-cell overlap,which requires laboratory professionals to spend a lot of time and effort to analyze and determine the cells in the smear.With the development of optical microscopy technology and computer technology,it is of great significance to use deep learning algorithms to identify a large number of cell images obtained and assist relevant personnel to complete the identification of cell states,which is of great significance to improve cell recognition efficiency and reduce manual analysis costs.In this project,a lightweight deep convolutional neural network ICARes2 Net and a Yolov3 detection algorithm based on dynamic loss are proposed for the classification and detection of smear cell images,and the corresponding application platform for automatic cell image classification and detection is developed.The main research contents of the paper are as follows:1.Aiming at the problems of low classification accuracy and poor real-time performance of Pap smear cervical cells,an improved coordinated attention module is proposed,and a lightweight deep convolutional neural network ICA-Res2 Net is designed by combining the new residual structure Res2 Net and the spatial pyramid pooling layer.Firstly,the cross-convolution between feature subblocks of the Res2 Net network is used to extract more fine-grained information in the feature layer.Then,the local area features are obtained through spatial pyramid pooling,furthermore,the improved lightweight attention module is introduced,which gives different weighted values to each pixel of the feature layer,strengthens the important detail features,and helps the network locate objects of interest.In addition,to effectively prevent the degradation of deep networks,the proposed ICA-Res2 Net network retains the hopin connection design of residual networks.Combined with the Softmax loss function and the center loss function,the network parameters are trained to improve classification accuracy.The experimental results on the cervical cell public dataset SIPa KMe D show that the ICA-Res2 Net network effectively improves the prediction accuracy and recognition efficiency of cervical cell images.2.Aiming at the problems of low detection accuracy of the Yolov3 model in blood smear blood cell image detection,a Yolov3 detection algorithm based on dynamic loss is proposed.On the one hand,considering that the semantic information obtained by the deep network lacks location message,and the lower layer network contains rich location information,the shallow information is denoised to reduce the noise impact,and further integrated with the deep semantic information to enrich the characteristics of the deep network.On the other hand,there is the influence of abnormal gradient values during training,so excessive gradient values are suppressed,thereby reducing the negative optimization of the network.Finally,Soft-NMS is added to solve the problem of missing detection caused by traditional NMS algorithms.The experimental results tested on the blood cell count dataset(BCCD)show that the proposed Yolov3-Dynamic Loss can detect various types of blood cells more accurately.3.Focusing on the smear cell classification and detection model in this paper,the application platform system of smear cell image classification and detection is designed and implemented in Java language and Python language.Through the modular design concept to process the classification and detection model,model management and image display functions are realized,and the classification and detection functions complete the corresponding tasks by calling the encapsulated Python program,providing prediction results to Web services,and realizing data sharing between different services.The results of image recognition are saved in the database in the form of entity data to form a result data set,to obtain the prediction results directly through the web terminal,further facilitate the use of relevant personnel,and realize the information management of smear cell images.