Detection Methods Of Cervical Abnormal Cells Based On Deep Learning

According to “Estimates of Incidence and Mortality of Cervical cancer in2018: a worldwide analysis”,cervical cancer is the fourth most common cancer of women.Due to large population and insufficient diagnosis and treatment resources,China has the most cervical cancer cases and deaths in the world.Currently,thin prep cytology test is an effective method for screening cervical cancer.This method mainly relies on manual diagnosis by pathologists,with great workload,high misdiagnosis rate,and missed diagnoses rate,leading to the high incidence of cervical cancer in China.In recent years,artificial intelligence has been combined with pathologists,and the intelligent assistant diagnosis system for cervical cancer has e merged.The system effectively reduces the workload of pathologists and improves the accuracy of diagnosis.As the core task of the system,the detection of abnormal cells is very important and challenging.These challenges include:（1）The image contents of cervical cytology smear samples are complex,which increases the difficulty of detecting abnormal cells.（2）Segmentation of cells is difficult due to image stainin g,cell adhesions and other factors.（3）Currently,machine learning algorithms measure images of all cells from all patients under the same criteria during training and recognition,reducing the accuracy of cell classification.（4）The cost of labelling cervical cytopathology images is high.The public datasets of cervical cells are few.The amount of data in different categories are unevenly distributed.To solve the above problems,the following methods are proposed in this paper.1.A target detection network called SER-DC YOLO is proposed to solve the problem of coarse detection for abnormal cervical cells in complex images.Firstly,the attention mechanism is added to Backbone of YOLO v5,which improves the feature saliency of the region of interest.Then,the normal convolutions are replaced with deformable convolutions in the spatial pyramid pooling layer,which improves the generalization of the network.Finally,the bounding box regression loss function is changed into the power exponential form,which improves the noise immunity of the network and the accuracy of the bounding box.It has been experimentally demonstrated that the SER-DC YOLO network reduces the rate of wrong and missed detection for abnormal cervical cells detection.2.A segmentation network called CE-GAN is proposed to solve the problem of accurate segmentation for cervical cell nuclei and cytoplasm.Firstly,cytoplasmic rough contours are extracted by the modified watershed algorithm in the coarse segmentation module.Secondly,cytoplasmic fine contours are generated using the Pix2 pix network in the guidance factor generation module.Thirdly,cytoplasmic fine contours are used in the fine segmentation module,constraining the training of the CE-Net model.Finally,cervical nucleus and cytoplasm are segmented by the CE-GAN model.It has been experimentally demonstrated that the CE-GAN network can improve the segmentation accuracy of cytoplasm by using boundary factor information.3.A method for fine classification of cervical abnorma l cells based on sample benchmark values is proposed to solve the problem of cervical abnormal cells reclassification.Firstly,the cervical cell images were cropped according to the predictions of the SE R-DC YOLO model to obtain Patch images.Then,normal squamous epithelial cells are layered.Baseline cells are selected from the mid-layer squamous epithelial cells based on TBS classification standard.The sample reference values are calculated using the CE-GAN segmentation results and diagnostic index formulas.Finally,the baseline coefficients for abnormal cervical cells are calculated.The feature classifier is trained.Abnormal cervical cells are reclassified.It has been experimentally demonstrated that this method can effectively exclude false positive cells and improve the detection accuracy of abnormal cervical cells.4.A semi-supervised classification method is proposed to solve the problem of classifying microbially infected cells when data are unbalanced.Firstly,the annotation data are filtered using the clarity evaluation function.Secondly,multiple deep learning classification models are trained using the filtered data.These models are assigned weights and then combined by specific strategies.Thirdly,the combined model is used to predict the class of unlabeled data and filter the "pseudo-labeled" data with high confidence in order to extend the training set.Finally,the training data are balanced by random downsampling and data augmentation operations.Classification models are further trained.It has been experimentally demonstrated that this method can effectively mine hidden associations in unlabe led data and improve the classification accuracy of microbially infected cells.In summary,to address the technical challenges in the research of intelligent assistant diagnosis system for cervical cancer,several specific solutions are proposed in this paper.These methods incorporate knowledges of pathologists and physician diagnosis,and cover technical areas such as target detection,image segmentation,and image classification.It has been experimentally demonstrated that our methods can effectively improve the performance of the assistant diagnostic system in detecting abnormal cells.