Research And Application Of Cervical Fluid Based Cell Detection Method Based On Single Domain And Multi Domain Scenarios

Cervical cancer is the most common malignant tumor in gynecology.In recent years,the incidence rate and mortality rate have shown an upward trend.Meanwhile,it has seriously troubled women’s life and health all over the world,and the rates are younger.Clinically,if cervical cancer can be found more early,the treatment effect can be better and the medical cost can be lower.Therefore,many countries are trying their best to popularize early screening of cervical cancer among women of the right age,in order to help women find cervical diseases and treat them in time at an early stage.Cervical fluid based cell test（TCT）is the best early detection method of cervical cancer determined by the international health organization.The advantage of TCT is that the detection rate of cervical abnormal cells is as high as 95%,which can effectively improve the screening accuracy.Besides,it reduces the missed diagnosis rate.However,TCT detection is highly dependent on the professional level of pathologists,which requires doctors to find abnormal cells through manual film reading.If doctors are inexperienced or overloaded,the misdiagnosis rate may be increased.Therefore,the deep learning based pathological image automatic cells detection method has become a research hotspot.Due to the wide coverage of cervical cancer screening,it can be carried out in large hospitals,physical examination centers,community health service centers and other institutions.The deep learning model of cell detection often needs to be applied to multiple medical institutions.Different medical institutions have different pathological section collection equipment and dyeing methods,and the accuracy,magnification and dyeing results of the images are different,resulting in different feature distribution of the dataset.In this situation,the application effect of the existing deep learning model in different hospitals is not well as expected,and there is a problem of domain shift.To solve the above problems,this paper first proposes a cervical fluid based cell detection model based on deformable convolution and attention mechanism,which improves the detection rate of cells in single domain scene;Then,a domain adaptive cervical fluid based cell detection model is further proposed to solve the domain shift problem,which improves the detection rate in the multi domain scenario;Finally,an online automatic labeling system for cervical fluid based pathological images is designed and implemented.The research content of this paper is mainly divided into the following three parts:（1）Detection of cervical fluid based cells based on deformable convolution and attention mechanismIn order to improve the accuracy of cell detection in single domain scene,a cervical fluid based cell detection method based on deformable convolution and attention mechanism is proposed.Firstly,aiming at the problem of large difference in the size of cells in cervical fluidbased pathological images,on one hand,the data enhancement technology is used to fill in cell samples of various sizes,so that the model can learn more regular cell contour features;On the other hand,the introduction of deformable convolution enables the model to further learn irregular cell contour features.Then,aiming at the interference of background categories in pathological images,spatial attention mechanism is used to strengthen the feature extraction ability of foreground categories,so as to reduce the interference of background categories to foreground categories in cell detection.Finally,experiments show that the proposed method can effectively improve the performance of cell detection in single domain scene,and can increase the m AP value by +1%,+4.8% and +3.9% respectively compared with the mainstream object detection algorithms.（2）Detection of cervical liquid based cells based on image-level and category-level domain adaptationIn order to improve the accuracy in multi domain scenes,a cell detection method based on image-level and category-level domain adaptation is proposed.Aiming at the problem of inconsistent feature distribution between source domain and target domain,this method adds image-level and category-level domain adaptive modules on the basis of one-stage object detection network retinanet.As a result,it realizes cross domain feature alignment and solves the problem of domain shift.In the image-level domain adaptation module,multi-scale image-level（i.e.global）features in the backbone network of source domain and target domain are extracted and input into the domain discrimination network for confrontation learning to realize global image feature space alignment;In the category-level domain adaptation module,the categorylevel（i.e.local）features are extracted by adjusting the classification subnet,which are clustered and compared for learning,so as to realize the local instance feature space alignment.Experiments show that the proposed method improves the cross-domain generalization effect of the model,and the m AP value on multiple datasets are increased by +12% and +6.8%respectively compared with the benchmark model.（3）An automatic labeling system for cervical fluid based pathological images based on webThe system uses MVC overall design pattern to separate the front and back ends.According to the practical needs of cervical fluid-based pathological image annotation,the standardized management of pathological image annotation is implemented through dataset management module,annotation/verfy module and prediction module.In the system,the cell prediction results of uploaded cervical fluid-based pathological images are implemented through the prediction module,so as to assist pathologists in disease diagnosis,which greatly improves the diagnosis efficiency.