Tumor Cell Image Recognition Based On Deep Learning

In recent years,deep learning develops rapidly,which is most widely used in speech recognition,image recognition,natural language processing and other areas.At the same time,deep learning has become the most popular research direction in the field of image processing and image recognition.At present,most of the classification and regression algorithms of tumor cell image belong to shallow learning.Shallow learning has limited ability to express complex functions and restricted to generalization ability of the complex classification problem.Shallow learning is also difficult to solve some of the more complex nature of signal processing problems,such as the human voice and natural image etc.But deep learning can represent the complex function with a few parameters through learning a deep nonlinear network,complex function approximation,and show strong ability of learning the essential characteristics focus on a sample data set.Therefore,this paper attempts to apply deep learning to tumor cell image classification and recognition,studying two kinds of deep learning model: convolutional neural network and deconvolutional networks.Convolutional neural networks has unique advantages in image processing among all the deep learning models.This paper firstly constructs a convolutional neural network aimed on the characteristics of tumor cell image with two convolution layers,two sampling layers and one layer fully connection layer.Then,the neural network model is improved and optimized by pre training and dropout technology to improve the robustness of the model.Then we use the expanded new image data set to train the CNN model.The model parameters are saved after the model converges,thus we can solve the problem of less image data.Finally,using the initialization parameters obtained in the above steps can greatly reduce the training time of the model and speed the whole model to accelerate the convergence when the model is trained the original tumor cell image data set.The designed CNN model and its classification accuracy for tumor cell image can reach 88%.The convolutional neural network has some problems in the process of learning,such as the parameter tuning process is not intuitive,the internal characteristics is opaque.To solve these problems,this paper uses the adaptive deconvolution network and analyzes the necessity and feasibility of this method.Through the model training and the inference of feature maps,the adaptive deconvolution network model is able to extract more abundant image features.Through the reconstruction algorithm the model can obtain and keep the reconstructed image similar to the input image.Moreover,ADN has faster training speed and inference.At the same time,combined with spatial pyramid matching algorithm,the proposed method uses SPM to train the extracted features by ADN,finally realization the tumor cell image classification.Through the simulation experiment,compared with CNN,the operation rate of the adaptive deconvolution network model was improved,the recognition accuracy can reach 91.7%.At last,the two kinds of methods of deep learning in this paper and tumor cell image of shallow learning classification methods were compared in simulation experiment,the recognition rate and speed of the two deep learning models increase to some extent,which can see that the study of deep learning image recognition of tumor cells is very promising.