Classification And Recognition Of Microscopic Cell Images Based On High-dimensional Feature Learning Of Small Samples

Cancer has become one of the most common and dangerous diseases in the world,and global morbidity and mortality are increasing every year.Early diagnosis of cancer is very important for effective prevention and treatment.At present,the gold standard for cancer detection is still pathological examination of histological slide,which provides an important basis for clinical treatment to a certain extent.With the rapid development of computer-aided diagnosis technology,researchers have proposed computer automatic analysis technology based on histopathology slides for comprehensive research and analysis of microscopic cell images.Due to uneven color,inconsistent illumination,and severe cell adhesion,accurately segmentation of cell nuclei is still a challenging task.The classification of microscopic cell images based on deep learning can automatically learn and find complex discriminative information from raw images,avoiding the error caused by manual diagnosis and cell nuclear segmentation.And it has good performance in the field of feature analysis.But deep learning usually requires a large number of labeled samples to train a reliable deep model.In order to solve the small sample problem of cell microscopic images,transfer learning is applied for image classification to avoid overfitting of the model.However,for the small sample set,transfer learning will generate high-dimensional high-level feature vectors,and the high-level features may contain redundant and irrelevant features,so that the fully connected layers are likely to be overtrained leading to degradation of its generalization performance.Aiming at the above problems,for the classification of microscopic images of mouse intestinal model,the microscopic image classification algorithm based on hybrid depth model and adaptive feature learning,and the microscopic image classification algorithm based on deep transfer network and rotational subspace ensemble learning are proposed in this thesis.They can solve the overfitting problem caused by the high-dimensional high-level features of small samples and improve the generalization ability of the model.The research contents of this thesis are:(1)Aiming at the problem that the high-dimensional high-level features of small samples easily lead to low generalization performance,the microscopic image classification algorithm based on hybrid depth model and adaptive feature learning is proposed.First,the Image Net dataset is used to pre-train the deep residual network,and the structure and parameters of the underlying convolutional layer and pooling layer are frozen for transferring to the target dataset.According to the characteristics of the mouse intestinal microscopic images,the new top structure is built on the pre-trained model and is fine-tuned to extract deep high-level features of the microscopic images.Then,the stacked sparse autoencoder is used to perform sparse dimensionality reduction on the high-level features and extract the sparse coding features of each hidden layer.According to the adaptive weight matching method,the best weight is assigned to the coding features of each layer,so as to achieve complementary advantages between features.Finally,the weighted features are cascaded and support vector machine is used to classify the fused features.(2)Aiming at the problem that the dimension of the high-level features is much larger than the number of samples,the microscopic image classification algorithm based on deep transfer network and rotational subspace ensemble learning is proposed.First,the transfer learning model is used to extract the high-level features from the raw microscopic cell images.Then,the rotational sample subspace sampling is used to increase the diversity between the high-level feature training sets,and the subspace projection method based on the manifold learning is used to perform nonlinear dimensionality reduction on the high-level features.Finally,the ensemble learning is used for classification based on the final dimensionality reduction features.This thesis mainly solves the overfitting problem of the high-dimensional high-level features of small samples based on deep transfer model,and enhances recognition ability of the model.The proposed method in this thesis provides an important tool for early cancer diagnosis,especially detecting the subtle change of the cell structure from pathological normal-appearing cell image,and has the good classification performance.