Classification Of Breast Lymph Node Metastasis Based On Deep Learning

Breast cancer is one of the highest cancers of incidence rate in women.It probably causes distant metastasis of cancer cells multi organ lesions and threaten the life of patients.Detection of lymph node metastasis based on pathological images is a key indicator of breast cancer staging,and correct staging decision is the premise and foundation for making correct treatment plan.At present,the detection of lymph node metastasis mainly relies on pathologist's artificial screening,which is time-consuming and laborious,and the diagnosis results of different doctors may be different.Deep learning technology can distinguish cancer assisted diagnosis according to pathological images,and has great potential for improving the accuracy of pathological diagnosis and optimizing the diagnosis process.In this thesis,the classification of breast lymphatic digital images metastases is studied based on deep learning.The convolutional neural network AFPCN based on the attention mechanism and less-parameter classifier is improved on the basis of the VGG(Visual Geometry Group)network.This article also optimizes image processing strategies and training strategies to improve model performance,and uses Grad-CAM to visualize the inside of the network to enhance the interpretability of classification results.Finally,this thesis successfully realized the automatic classification of metastatic cancer based on the digital pathological images of breast lymph nodes.The main work of this thesis is as follows:(1)This color invariance algorithm is combined with multi-scale image processing strategies to solve the coloring chromatic aberration and overfitting.Color Constant Algorithm can normalize the images of metastatic cancer with serious color differences,and reduce the negative influence of staining reagents difference.In addition,multiscale feature extraction strategy is combined with various image enhancement methods to increase the diversity of the data set and delay the timing of network overfitting.(2)It is proposed to use the DCSCN super-resolution network to solve the problem of large resolution differences in digital pathological images of different data sources in practical applications.After the resolution,the image is increased by 2 times,which provides more effective feature information for the network and improves the network generalization performance.(3)This thesis proposes an AFPCN network which is suitable for the classification of pathological slices of lymphatic metastases.The network has the ability to automatically screen effective features by introducing attention mechanism.Combined with the training method of spontaneous thermal restart,the final accuracy rate is 4.37% higher than that of the original network,and the parameter quantity is reduced by 68%,and the AUC value reaches 0.9928.The method has achieved the expected result on the Patch Camelyon pathological image dataset of breast lymphatic metastasis.The accuracy of the single model is as high as 0.9603 on the test set,which is 1.39% higher than the latest multi-model integration method.Automatic identification of breast lymphatic metastasis by computer can provide accurate and objective evaluation basis for pathologists,and save time and effort.At the same time,the accuracy of the method in this thesis zero sample training on the lung cancer pathology database Lung HP is 0.8212,which indicates that the system has strong generalization performance on other cancer classification tasks based on pathological slices.