Research On Cell Image Recognition Of Cervical Lesions Based On Convolutional Neural Network

Combined with computer technology and pathological theory,the automatic classification and recognition of cervical lesion cells has important practical value for the diagnosis of cervical cancer.Traditional methods of cervical lesion cell recognition and classification need to extract features manually,which will lead to subjective errors,resulting in poor recognition results.In this paper,cervical cell images are classified into four categories.Combined with convolution neural network,the accurate recognition and classification of cervical cell images are realized.The main work is as follows:(1)In this paper,a convolution neural network S-CNN model with parallel convolution layer was designed.The model consists of seven convolution layers,four pooling layers,two full connection layers and output layers.The convolution layers and pooling layers are connected alternately.By adding parallel convolution layers,the output feature images are superimposed to obtain more abundant cell image features.The superimposed feature images are input to the next convolution layer,and finally classified by the Softmax classifier through two full connection layers.The batch normalization algorithm optimizes the model and speeds up the network training.The experimental results show that compared with Alex Net and VGG16 convolutional neural network models,the S-CNN model designed in this paper can more accurately learn the characteristics of cervical lesion cell images and effectively reduce the recognition and classification error rate.(2)An improved cross-entropy cost function algorithm based on residual network was proposed.According to the real label and predictive label in cervical cell image recognition,the weight matrix is established and applied to cross-entropy cost function to improve the algorithm.In the recognition and classification of cervical cell images,false negative recognition and classification errors of cervical lesion cells,i.e.false negative rate,will directly affect the diagnosis of cervical cancer.In order to reduce the false negative rate of cervical cell image recognition and classification,the weight matrix elements are classified into different levels,and different levels are given different weights.The experimental results show that the improved cross-entropy cost function algorithm can effectively reduce the false negative rate of cervical lesion cell recognition and classification,thus reducing the overall recognition and classification error rate.