Research On Cervical Cell Detection Method In Liquid-based Cytology Images

Liquid-based cytology is one of the most advanced detection techniques for cervical cancer and can effectively prevent cervical cancer.The existing mainstream methods rely on manual screening of liquid-based cytology images,but manual screening is time-consuming and inefficient.Faced with a large number of liquid-based cytology images,people have an increasingly strong demand for computer-assisted cervical cell detection tools.The detection performance of traditional multi-stage cervical cell detection methods is limited by the quality of image segmentation and feature extraction,which is not conducive to the application of large-scale practical scenes.Existing object detection technologies based on convolutional neural networks are developing rapidly in cervical cell detection tasks.However,these methods lack specific design for the characteristics of cervical cells,and their detection performance needs to be improved.Therefore,this paper will mainly optimize the existing object detection techniques to make them suitable for liquid-based cytology images of cervical cells.In addition,from the perspective of practical application,this paper will also study the development and utilization of a large number of unlabeled cytology image data based on semi-supervised object detection technology.In this paper,we first study the supervised cervical cell detection algorithm.In view of the irregular cell morphology of the cervix,this paper describes in detail the problem of misalignment in label assignment using existing object detection techniques.To this end,this paper alleviates the impact of label assignment misalignment from two perspectives.Firstly,this paper directly deals with the label assignment algorithm of object detection,and proposes the dynamic mean-standard deviation label assignment algorithm.The discriminative scores of feature points are mainly constructed by loss function,and the original label assignment based on intersection over union is changed to discriminative scores.Secondly,this paper indirectly starts from the optimization of features to improve the perception ability of feature points for context information,so that they can still maintain rich semantic information in the case of unreasonable label assignment to predict correct results.In the process of experiment,this paper established the cervical cell data set,the experimental results show that the proposed supervised object detection model effectively improve the performance in the cervical cell detection task,more than the existing general object detection method and the method specifically designed for cervical cell detection,the highest can reach 57.6 % m AP,And can achieve the best trade-off between speed and accuracy.In order to further develop and utilize unlabeled data,this paper studies semi-supervised cervical cell detection methods,including semi-supervised self-training object detection and semi-supervised end-to-end object detection.The training process of the semi-supervised self-training method is complicated,and the existing semi-supervised end-to-end object detection method has the irrationality of obtaining pseudo-labels.Therefore,in this paper,based on the semi-supervised end-to-end mode,the structure of the basic detection model is not changed,and the classification branch is re-endowed with the semantics of quality estimation,so that it can select higher quality pseudo-labels.Experimental results show that the proposed method can effectively develop and utilize unlabeled information,and exceeds the existing general semi-supervised object detection method and semi-supervised detection method for cervical cell detection task in the presence of 50% or less labeled data.This study aims to improve the performance of object detection technology in cervical cell detection task.From the perspective of practical application,considering the irregular morphology of cervical cells and the large amount of unlabeled data to be developed,this paper specifically studied the application of object detection technology in cervical cell detection task under supervised learning and semi-supervised learning respectively.This study is for liquid-based cytology examination,further protect the health of women,play a certain positive role.