Cervical Cell Nuclear Segmentation And Recognition Method Based On Convolutional Neural Network

Cervical cancer is the fourth most common cancer among women worldwide.The current effective diagnostic method for cervical cancer is liquid-based thin-layer cytology.But this requires doctors to look for cancer cells in a large number of cells under the microscope.Its workload is huge,the rate of misdiagnosis is high,and there is a serious shortage of pathologists in China.Therefore,an intelligent auxiliary diagnosis system is urgently needed.The segmentation and identification of the nucleus are the two key steps to determine whether this system is effective.However,the images collected by the microscope will have some uneven illumination,complex background,and different shades of staining.There will also be some cell debris and garbage in the collected images.And the image seen by the doctor under the microscope has multiple levels,and the focal points of the cells at each level may be different,but the image collected in this article is a 2D planar image,which wil l cause part of the collected cell image to be out of focus.Clear,and there are a large number of overlapping cells.At the same time,in the classification process of cell images,normal cells and cancer cells coexist,the difference between them is sma ll,and it is difficult to identify.In response to the above problems,this paper first proposes a cervical cell nuclear segmentation method based on optimized maximally stable extremal regions(Maximally Stable Extremal Regions,MSER)algorithm,which effectively eliminates the influence of complex background on cell nuclear segmentation.Then,a method of segmentation of overlapping nuclei based on U-Net is proposed to solve the problem of segmentation of nuclei in overlapping cells.Finally,for the classification of cells after segmentation,a cell classification method based on deconstruction-reconstruction module is proposed.In this paper,we take cervical cell images as the research object.Combining the two steps of cell segmentation and cell classification after segmentation,the screening of lesioned cervical cell nuclei is realized.The specific work is as follows:1.A method of segmenting cervical nucleus based on optimized MSER algorithm was proposed.This method first converts the image to the HSV(Hue,Saturation,Value)color space.Then,after weighting the S and V channels,the weighted image is processed using the optimized MSER algorithm to obtain a coarsely segmented region with uniform gray values.Then the parameter adaptive threshold segmentation method is used for fine segmentation.Finally,the artificial neural network is trained by extracting the features of the nuclei after the fine segmentation to determine whether the result obtained after the segmentation is the nuclei.2.An overlapping cell segmentation algorithm based on U-Net model was proposed.On the network structure,we reduce the depth of U-Net,use two down-sampling and two up-sampling operations in the encoder and decoder,respectively,to improve the efficiency of network segmentation.Attention mechanism is introduced to focus on the boundary pixels of overlapping cells.At the same time,the 3 × 3 convolution in the network is replaced by an asymmetric convolution to improve the accuracy of model segmentation while ensur ing efficiency.3.A classification method of cervical cancer cells based on the deconstruction and reconstruction module was proposed.This method first uses the region obfuscation mechanism to scramble the original cell image,and then inputs the scrambled cell image and the original image into the convolutional neural network for training.The obtained feature vector is used for classification learning,adversarial learning,reconstruction learning.We get the final cancer cell classification network.Experiments show that the method proposed in this paper has a good segmentation effect on discrete and overlapping nuclei.They improved segmentation efficiency,which can basically meet the needs of nuclear segmentation in the cervical cancer auxiliary diagnosis system.At the same time,the classification accuracy of the convolutional neural network based on the deconstruction and reconstruction module can reach more than 95%.This article can provide a reference for the computer-aided diagnosis of cervical cell images.