Study On Diversity In Microscopic Hyperspectral Pathological Image Based On Deep Neural Network

Pathological image recognition based on microscopic hyperspectral imaging technology is a new and important research direction of digital pathological analysis.Its significance lies in solving the problem of limited information in traditional microscopic imaging,thereby effectively improving the accuracy of pathological analysis.Based on the current microscopic hyperspectral imaging technology,the wavelength range of pathological analysis is mainly concentrated in visible light and near-infrared.As the scope of pathological analysis becomes wider and wider,the diversity of pathological tissues and the complexity of structural features greatly increase the difficulty and complexity of pathological identification.The development of deep learning solves the problems of traditional methods such as strong subjectivity of feature construction and low fusion efficiency,so as to make more efficient use of microscopic hyperspectral features.At present,deep learning-based microscopic hyperspectral recognition algorithms mainly use low-dimensional,shallow convolutional networks,resulting in sole feature,which does not make full use of the advantages of microscopic hyperspectral.At the same time,the model is relatively simple and insufficient to solve complex pathologies.analyse problem.The use of deep convolutional networks can solve these problems,and many excellent methods have emerged in the field of medical image processing.However,most of these algorithms design network models for specific analysis tasks of a certain disease,and focus on segmentation tasks,and lack the same model architecture for segmentation,classification,and detection tasks.Meanwhile,most of the end-to-end models based on deep convolution are mainly used in threedimensional imaging modes such as CT and MRI.The high variability and complexity of pathological images are not fully considered,so how to establish end-to-end deep convolutional network suitable for microscopic hyperspectral pathology image is an important problem to be solved urgently.Therefore,this thesis comprehensively analyzes multi-granular feature fusion,hierarchical feature enhancement,and multi-channel feature combination methods,and establishes a recognition architecture based on deep convolutional networks and multi-feature guidance,which improves the recognition accuracy of highly variable cells and tissues.Research is carried out on key issues such as segmentation,classification and detection in microscopic hyperspectral pathology recognition,and the core architecture has been applied to and verified on skin melanoma,white blood cell and blood cell images.The main work is as follows:(1)Focusing on the application of deep convolutional neural networks to solve the problems of the structure and deformation diversity of cells and tissues in microscopic hyperspectral images,a multi-channel and multi-level guided end to end pathological image recognition network architecture is proposed.By comparing the characteristics of different hierarchical feature learning,fusion,and enhancement methods,and a multi-channel guided three-dimensional convolutional neural network is designed.In the experiments of microscopic hyperspectral tissue segmentation,cell classification and detection,the proposed method has achieved better results than the single feature based neural network recognition method.(2)Aiming at the problem of exploring the mechanism of diverse feature extraction and realizing multi-feature guided fusion segmentation,a segmentation framework of threedimensional fully convolutional neural network with multi-granular feature fusion is proposed.We analyze different fusion methods for adjusting low-level and high-level semantic features,and construct a multi-feature guided segmentation network based on dual-path feature fusion and residual feature fusion methods.With the help of an adaptive loss function composed of weighted loss and regional loss,the proposed segmentation method has achieved higher accuracy compared with the single feature-guided fusion segmentation method.(3)Aiming at the key issue of how to realize the multi-feature guided classification under the diverse attention mechanism,a classification method based on the hierarchical attention mechanism of a three-dimensional convolutional neural network is proposed.We study the image augmentation and regularization method to sovle cell diversity,and design a multifeature guided classification network based on grouped convolution and attention mechanism.Through gradient backpropagation and visualization technology,the correlation between the spectral features contained in the microscopic hyperspectral image and the differentiation of cell types is explored.(4)Aiming at the key issue of exploring how the multi-feature mechanism can improve microscopic hyperspectral cell detection,a multi-feature guided three-dimensional single-stage cell detection network is proposed.We study different anchor mechanism and regression box fusion methods,and propose a cell detection network based on multi-task joint anchor mechanism and weighted regression fusion method,which solves the missed detection and low regression accuracy caused by differences in the number and morphology of cells.In experiment,the proposed method has achieved better real-time performance and robustness.