Research On Deep Learning Methods For Abnormal Cervical Cell Detection

Cervical cancer is the fourth cancer that threatens women’s health.According to the WHO 2018 Global Cancer Observation Database,there were approximately570,000 cervical cancer cases and 311,000 deaths worldwide in 2018,with a mortality rate of 54.56%.Screening for cervical cancer in advance can enable the cancer to be detected early.This helps to treat cervical cancer and reduce mortality.Thinprep Cytologic Test(TCT)has become the main method of screening for cervical cancer.Since there are thousands of cells on a single slice,pathologists with rich experience will encounter many missed and misdetected cases when reading the film.Automatic reading technology can assist doctors in completing the cervical cancer screening process,which not only reduces the missed detection rate and false detection rate,but also improves the detection efficiency.Precise screening of abnormal cervical cells is a key technology for automatic image reading.Commonly used methods for screening abnormal cervical cells include traditional detection methods,convolutional neural networks and target detection methods.These methods can achieve a certain screening accuracy,but there are still some detection problems that have not been resolved.Generally,in the same image,the size of abnormal cell nuclei is quite different from that of normal cells,but the current target detection method does not maintain the uniform size characteristics of cervical cells in the classification stage.Normal cervical cells and abnormal cervical cells belong to the subcategories of cervical cells,and the similarity between them is relatively large,so the cervical cell classification task belongs to fine-grained classification.For a single cell,it is not easy to judge whether it is abnormal or not.It is easy to judge its category after comparing with surrounding cells on the same image.The current classification method generally only inputs one cell image at a time,and does not use comparison with surrounding cells.This article focuses on the use of deep learning methods to solve the above problems,and improves the classification accuracy of cervical cells from three aspects: target detection,fine-grained classification and multiple input networks.The specific work content includes the following three parts:1.Designed a deep neural network MACD R-CNN based on Mask R-CNN,which solves the problem that Mask R-CNN does not uniformly maintain the characteristics of cell size and shape in the classification branch.First,in the classification branch of Mask R-CNN,it generates feature maps of the same size as the input from Ro Is of different sizes.The nuclei of this part of the feature map will be deformed to varying degrees.We designed a fixed proposal module to generate a fixed-size nuclear feature map so that new nuclear features can be used for classification.Then we use the attention mechanism to fuse the original Ro I and fixedsize Ro I features.Finally,we increased the depth of the convolutional layer to further improve the accuracy of cell classification.Experiments show that MACD R-CNN can effectively improve the performance of abnormal cell detection.2.The weakly supervised classification model MNTNet is proposed to solve the difficult classification problem due to the diverse changes of cervical cells,the complex environment,and the greater similarity between abnormal cells and normal cells.The performance of current cell classification methods is still not ideal,and MNTNet introduces the idea of fine-grained classification to have better effects on the above problems.We adopt the anti-binary tree design method.First,the spatial attention mechanism and channel attention mechanism are used to learn different features on the left and right child nodes,and then the left and right child features are merged through the merge operation of the parent node,and finally the root node is used to realize the cell classification.The experimental results on the Herlev data set and SIPAKMe D data set show that the accuracy of our proposed MNTNet method surpasses the current advanced methods.3.The MINDDR-CNN-Shortcut multi-input cell classification network is proposed,which solves the problem that ordinary classification networks cannot use the feature comparison and connection of cervical cells in the entire image.The MINDDR-CNN-Shortcut network is a multi-input single-task network used to detect abnormal cervical cells.The main contributions of the MINDDR-CNN-Shortcut are:firstly,adapt to the simultaneous input of multiple images;secondly,increase the CBAM-NDDR module to make the feature fusion of the sub-networks more smooth;finally,connect the feature maps output by each sub-network CBAM-NDDR together,And further integrate the characteristics of each sub-network.Experiments show that our proposed MINDDR-CNN-Shortcut network can learn the characteristics of multiple cells in an image and has a higher accuracy rate than the single network model.Deep neural network is an important tool to realize computer-aided automatic diagnosis technology.When using deep learning technology to solve cervical cell classification problems,it is not only necessary to understand the types and characteristics of cervical cells,but also to combine certain pathological knowledge.In this paper,by analyzing the existing problems of the current abnormal cervical cell detection methods,combining deep learning technology with cervical cell characteristics and pathological knowledge,some new detection methods are proposed.These methods can not only improve the detection efficiency of cervical cancer cells but also improve the detection accuracy.