Classification Algorithm Of Microscopic Cell Images Based On Deep Transfer Learning

Cervical cancer has become one of the major diseases affecting women’s health,and the incidence is increasing.Early diagnosis is the key to prevent and cure cervical cancer.Pathological examination based on cervical pap smear images is one of the main methods for the diagnosis of cervical cancer.With the continuous development and innovation of computer-aided diagnosis technology,researchers have proposed computer-aided diagnosis methods for automatic analysis of cervical pap smear images.The computer-aided diagnosis methods based on manual features has some problems,such as poor contrast,uneven staining and overlapping cell nuclei,which make it difficult to accurately segment the cervical pap smear images.The classification of cervical pap smear images based on deep learning can learn and extract effective recognition information from the original data,and avoid errors caused by cell segmentation.However,due to the large number of parameters,deep learning requires a large number of labeled samples to establish a stable model.For the small sample problem of cervical pap smear images,transfer learning is introduced into image classification.However,the classification of cervical pap smear images based on transfer learning faces the overfitting problem caused by high dimension features of small number of samples and the poor generalization ability of the transfer model,which limit the accuracy and stability of the model.In order to solve the above problems,two classification algorithms of cervical pap smear images based on deep transfer learning are proposed in this paper.These two algorithms solve the overfitting problem and improve the generation ability of the model by reducing the model parameters and the feature dimension,respectively.The research work of this paper mainly includes:(1)A method of cervical pap smear images classification based on transfer learning and adaptive pruning is proposed.Firstly,the source data set is used to train the pre-trained network,then the model is reconstructed according to the characteristics of cervical pap smear images,and then the parameters of convolution layers and pooling layers in the pre-trained model are migrated to the reconstruction model.On this basis,the method prunes the convolution kernels and neurons of the model.Considering the interaction between the source data set and the target data set,we utilize the pruning criterion combining the convolution kernel weight value and the output feature maps.Through the pruning and optimization for the model after transfer learning,the optimal classification model of cervical pap smear images is finally obtained.(2)A classification method of cervical pap smear images based on hybrid depth network structure is proposed.Firstly,the convolution neural network is used to extract the high-level features of cervical pap smear images directly.Then,according to the characteristics of the cervical pap smear image data set,the manifold learning method is used to reduce the dimensionality of the high-level features of the cervical pap smear images,remove the interference noise in the original features,and finally classify the features using support vector machine.The research work in this paper solves the problem of overfitting in the classification of cervical pap smear images based on transfer learning,and improves the generalization performance of the model.It provides a new solution for the computer-aided diagnosis of cervical cancer,and has certain theoretical significance and clinical application value.