| Global cancer epidemiology data for 2022 show that breast cancer has successfully surpassed lung cancer to become the most common cancer in the world.However,early detection and timely treatment of breast cancer will greatly reduce the mortality rate of patients.Mammograms,ultrasounds,MRIs and CT scans are currently used to detect breast cancer,but they are expensive and prone to misdiagnosis.The advent of histopathological imaging of breast biopsy has brought benefits to groups at high risk for breast cancer.The medical images,obtained through the microscopic technique of biopsy tissue,help experts to accurately identify subtypes of breast cancer while also allowing them to observe tissue characteristics on a cellular basis to pinpoint areas of disease.Therefore,histopathological image diagnosis is often regarded as the "gold standard" for breast cancer detection.At present,the difficulty of medical image analysis lies in the need of experienced histopathologists,but the training of these talents requires a lot of time and resources.To make matters worse,imaging diagnoses can easily be misjudged by doctors.With the improvement of image processing technology and the rapid development of machine learning,people are constantly trying to use computer technology to improve the accuracy and efficiency of diagnosis.At present,a variety of deep learning algorithms are applied to classification and recognition of breast cancer histopathological images.However,these methods not only have a large number of parameters,but also cannot obtain global features well,resulting in the loss of a large amount of effective information in the process of model learning.Tensor is a high-level generalization of the mathematical form of vector and matrix,which can be used as data storage.When used for image classification,it can reduce the cost of data storage while maintaining the spatial correlation between image pixels.Tensor network is a kind of network formed by the contraction of several tensors according to specific rules.It is possible to decompose higher-order tensors into related lower-order core tensors to reduce the parameters.Compared with the traditional tensor network,the quantum tensor network originated from the multi-body quantum physics maps the input data into the Hilbert space,which can well maintain the internal structure information of the high-dimensional spatial data and solve the problem that the traditional two-dimensional matrix is difficult to describe the relationship between the data.Therefore,the quantum tensor network model may become a breakthrough in the classification of breast cancer medical images.The main work completed in this thesis is as follows:(1)This thesis proposed Quantum Tensor-augmented Convolutional Representations Network(Con Ten Net)based on Convolutional Neural Network(CNN)and Quantum Tensor Network(QTN).The Con Ten Net is used for binary classification of the histopathological images of breast cancer.The QTN is applied to the classification of breast cancer histopathological images.The number of model parameters is reduced by the properties of parameter compression of the QTN.Information connection between adjacent regions is realized by extracting data from adjacent regions and transmitting information between layers through QTN,so as to obtain deeper global features.At the same time,it also lays a foundation for the further combination of QTN and classical model.In order to ensure the balance of training data set,the data enhancement method is used to enhance the benign image and reduce the interference of color to the experiment.The color normalization of the original image is carried out.In this thesis,comparative experiments are carried out on images with different magnifications from two perspectives of image-level and patient-level.Experiments show that the accuracy of this model can reach about 99.06% under 40 x of the original image-level,and about 92.98%under 40 x of the patient-level.The accuracy rate of color normalized image was 97.56%under 40 x of image-level,and the accuracy rate of patient-level was 91.89% under 40 x.Finally,the Grad-CAM method is used to prove that although the accuracy of color normalization decreases,the method is important and necessary,and makes the binary fusion model reliable and interpretable.(2)A fusion model based on Con Ten Net and texture feature extraction is proposed for multi-classification of histopathological images of breast cancer.The characteristics of parameter compression and feature extraction of Con Ten Net model are migrated to the Con Ten Net model.Meanwhile,in view of the complexity of multi-classification of histopathological images of breast cancer,the traditional texture feature extraction methods,namely Gray-Level Co-occurrence Matrix(GLCM),Local Binary Pattern(LBP)and Gabor filtering,are introduced.The GLCM method was fused with LBP and Gabor filtering method respectively,so as to obtain the local cell shape characteristics of breast cancer histopathological images in space and different scales and directions.Finally,the data is input into the full-connection layer for classification through feature fusion.The experimental results show that the accuracy of the fusion model can reach 97.20% under100 x of the original image-level,and 97.07% under 100 x of the patient-level.The accuracy of color normalized image can reach about 95.67% under 40 x of the image-level,and the accuracy of patient-level can reach about 95.34% under 40 x,both of which are better than the common comparison models at present.Finally,the confusion matrix analysis of the experimental results shows that the model has a good classification of eightegories,showing that the model has a strong resolution and good performance. |
PDF Full Text Request