Detection Of Cervical Cancer Cell Based On Strong Characteristic CNN-SVM

With the continuous development of medical image processing and pattern recognition technology,cervical cancer cell detection technology based on machine vision emerges as the times require.This technology is expected to replace traditional manual screening methods to solve the problems of low efficiency,being greatly influenced by subjective factors and low accuracy in manual screening methods.Because the technology is in the initial stage,there is a general problem of low intelligence of equipment,low accuracy of recognition,poor sensitivity and specificity of classification.Aiming at this problem,this paper studies a set of automatic detection system for cervical cancer cells on the basis of existing research techniques.In this system,a microscopic imaging device with the improved focus algorithm is used to photograph the smear of the cervix exfoliated cells,and then the lymphocyte image and the epithelial cell image are classified by the random forest classifier optimized by the parameter.Finally,a strong feature CNN-SVM classifier is introduced to classify normal epithelial cells and cancerous epithelial cells.The specific work of this article is as follows:(1)In the design of hardware platform,X and Y axis intelligent control modules are added to enable the device to switch image horizons automatically.In the auto focus algorithm,the focusing window selection strategy is improved.The focusing window is selected automatically according to the density of the target within the field of view.Then,the image sharpness is evaluated by the multi gradient definition evaluation function.Finally,the focal plane is searched rapidly by hill climbing method.(2)Because of a large number of irrelevant lymphocytes in the smear of cervical exfoliative cells,a random forest classifier model based on artificial fish swarm optimization is studied in order to classify the epithelial cell images and lymphocyte images quickly.This classifier makes use of the artificial fish swarm algorithm to select the scale of Random Forest and characteristic subset,to ensure the validity and diversity of sample subspace and characteristic subspace,improve the accuracy and generalization ability of classifier.(3)To solve the problem of low accuracy,sensitivity and specificity of classifying epithelial cells and normal epithelial cells,a CNN-SVM classifier model based on strong features is studied in this paper.In this model,the feature extracted automatically from the hidden layer of CNN network is sequentially fused with the extracted strong features.Then the features are inputted to support vector machine for classification.The classification model adds strong feature paths to guarantee the important feature of the sample to be learned by the classifier.The classifier improves the accuracy,sensitivity and specificity of recognition.Experimental results show that the improved auto focusing algorithm has higher efficiency and better focusing effect.The OOB error of the Random Forest model after parameter optimization is as low as 0.0314,which is more generalization than other classification methods.The accuracy of CNN-SVM classifier based on strong feature is 94.92%.Besides,the model ensures the sensitivity and specificity of the classification.