Pathological Graded Diagnosis Method Of Cervical Cancer Cells Based On Deep Learning

Health issues have always been the focus of human attention.Although the development of modern technology has promoted the improvement of China’s medical health,cancer,as the number one killer of human health,is still one of the most difficult to treat diseases.Among women’s common cancers,cervical cancer ranks second in both morbidity and mortality,which ser iously threatens women’s lives and health.Nowadays,the diagnosis method of cervical cancer is for professional doctors to look for cancer cells from field to field under a microscope.The workload is heavy and it is easy to miss and misdiagnose visual fatigue.In recent years,artificial intelligence has entered the medical field,and it is expected that it can accelerate the progress of smart medicine and provide more assistance to doctors.Therefore,the rapid and accurate mass screening of cervical cancer using deep learning methods has become a hot research topic.However,in order to realize the automatic identification and grading of cervical cancer cells,there are the following technical difficulties: First,the content of cervical cell images is complicated.Then,deep learning is not sensitive to object size.The structure of cancer cells and normal cells are very similar.Finally,grading data is difficult to obtain and requires professional doctors to mark it.Not only is it expensive,but the distribution between categories is extremely uneven.In order to solve the above problems,this article proposes the following solutions:1.A method for cervical cell nuclei classification based on convolutional neural network is proposed,which solves the problem of identifying cervical cancer cells in a complex background.This method firstly improves the U-Net model,improves the efficiency and accuracy of network segmentation,segments the nucleus of the cell image,then uses active learning to subdivi de the nucleus,and uses the Res Ne St network to extract image features,and then introduces the doctor’s judgment index as artificial features,Combining the two types of features to determine the type of nucleus,reducing the impact of other cells.Finally,cervical cancer cells were screened according to TBS diagnostic indicators.Experiments show that this article can analyze the components of the cell image well,and classify each cell nucleus.The classification model introduced into the doctor’s experience is more accurate.2.A method for identifying cancer cells based on YOLO is proposed,which solves the problem that the structure of cancer cells and normal cells is similar,the size and texture of the nucleus is very small,and the deep network is not sensitive to the size of the object.The method in this article first modifies the convolution of the YOLOv5 model.First,the symmetric convolution of the backbone is modified to an asymmetric convolution to obtain better feature expression without additional overhead;then dynamic convolution is introduced to make the convolutional network based on Input the cell image to dynamically modify the convolution parameters,and extract the features with the appropriate convolution kernel size.Then add the attention mechanism behind the backbone of the last two layers,pay attention to the interactive relationship between the cell image channels,and pay attention to the size difference of adjacent cells.Finally,select part of the patient image to segment the cell nucleus and count the area and other information,use the doctor’s diagnosis experience to design a score formula,calculate the score of the cancer cell,and rank and recommend it to the doctor.Experiments show that this method not only improves the accuracy of the cervical cell image data set by 6%,but also improves the accuracy of the public data set COCO by 2% compared with the original network.3.A grading method for cervical cancer cells based on long-tail data is proposed.The problem of difficulty in obtaining hierarchical data and unbalanced distribution of various categories is solved.This method first divides two branches to learn the general features of the sample and the tail sample features,and uses asymmetric convolution to extract the features,and then adjusts the cumulative learning weight function to find a more suitable function.Finally,the rules for judging the level of positive patients are formulated based on the doctor’s experience.Experiments show that this method can not only achieve considerable accuracy on the cell grading data set,but also improves by 1.22 percentage points on the public data set CIFAR.In summary,this article proposes specific solutions to the problems in the screening and grading of cervical cancer cells.Experiments show that this method can well detect and grade cervical cancer cells,and integrate the doctor’s diagnostic experience and TBS diagnostic indicators throughout the entire diagnostic process.In addition,this article also examines the model with the patient as the unit to effectively solve the actual diagnosis problem and realize the automatic detection and grading of cervical cancer cells.