Research Of Cell Image Segmentation And Classification Methods Based On Convolution Neural Network

At present,the main way of diagnosis of cervical cancer is Pap test by cytology microscopic,medical personnel through pathological knowledge and their own experience in reading cell smear for diagnosis.Manualreading of cell smears is a depletion of time and labor,and the diagnostic accuracy is low.Convolution Neural Network is used to analyze the image of cervical cells,so as to realize automatic film reading by computer,so as to reduce the medical workload.Firstly,a NS-Net network is proposed for the problem of cell nuclear segmentation.The encode-decode structure of the traditional segmentation network is continued.Deep features and shallow features are integrated together by skip connection;Atrous Spatial Pyramid Pooling can obtain the contextual multi-scale information;designing A-type multi-scale bottleneck layer and placing it on the shrinking path,splicing features of different levels to obtain more information;designing a B-type multi-scale bottleneck layer and placing it on the expansion path,extracting fusion features ulteriorly;introducing an attention mechanism to improve visual information processing efficiency and accuracy.The experimental results show that NS-Net can achieve fine segmentation of cell nuclei with ideal results.Secondly,for the problem of cell nuclear classification,a residual network combined with depth separable convolution is proposed.A deeply separable residual structure is designed.The structure is divided into three paths,namely,a normal convolution path,a depth separable convolution path,and a trainable skip connection path.On the basis of the traditional residual block,combined with the depth separable convolution operation,the network’s independent analysis ability of features is enhanced.With Res Net50 as the basic framework,three arrangements are designed: layer-by-layer arrangement,polarization arrangement,and alternate arrangement.The alternate depth and separable residual blocks can realize multi-scale feature mining.Experiments show that the network has a good ability to distinguish between normal cell nuclei and cancer cell nuclei.Finally,NS-Net and the depth separable residual network are encapsulated as an application,the graphical user interface is realized by Py Qt in Python environment.By triggering the corresponding buttons,the readers of cell smear can obtain the results of cell nucleus segmentation,cell positioning,and cell classification of a specific picture.The reader of cell smear can obtain diagnostic references from the results of the picture processing,thereby reducing the doctor’s work threshold and reducing the doctor’s workload.