Histopathological Image Classification Of Breast Cancer Based On Deep Learning

As one of the most lethal and prevalent cancers in women worldwide,histopathological analysis is the main method for breast cancer diagnosis.In this paper,we propose two deep learning-based methods for classifying breast cancer histopathology images based on the Break His dataset of breast cancer histopathology images,and the experimental comparison analysis can verify the effectiveness of this paper’s methods for improving the recognition accuracy of benign and malignant classification of breast cancer pathology images.The main contents of the paper are as follows.(1)To address the problem that single classifier and integrated learning classifier models have a limited observation domain and tend to fall into local optima,a classifier model based on joint training is proposed.First,we combine multiple classifiers SVM,KNN,decision tree,random forest and plain Bayes to expand the hyperparametric observation perceptual domain in order to have less lossy estimation points,then use Bayesian optimization to find the best observation,find the direction of the least lossy estimation point according to the best observation,continue to find the least estimation point in this direction in the next iteration,and update the estimation point according to the found estimation point if it makes the loss lower The Gaussian process fitting function is updated according to the found estimation point.If the optimisation gradient becomes negative at a certain value,the local optimum is reached and the next point where the gradient becomes positive is found by adding a random number of perturbations to jump out of the local optimum until the end of the iteration.The optimisation hyperparameters are iteratively optimised according to the best estimation point and thus a classifier with better fitting performance is jointly trained.The experimental results of the single classifier and the jointly trained classifier proposed in this paper show that the accuracy of the jointly trained classifier is improved for different magnifications of images,with 99.67%,98.08%,99.01% and 96.34% obtained at 40×,100×,200× and 400× respectively.The sensitivity of the four magnifications was 99.51%,97.91%,99.28% and 93.24% respectively,which largely prevented malignant breast cancers from being missed.F1＿score of 99.22%,97.16%,98.84% and 95.94% respectively also illustrated the high accuracy of malignant samples while not being missed.(2)Based on the jointly trained classifier model,a super-resolution and feature fusion-based classification method for breast cancer pathology images is proposed using low-dimensional features,considering that the extraction of convolutional high-dimensional features will increase the computational time and computational effort,to improve the recognition accuracy while reducing the computational effort.Firstly,a super-resolution generative adversarial network model is used to process the breast cancer pathology images to obtain clearer and more realistic super-resolution images with more texture details.The R and B channels of the RGB image and the V channel of the HSV image are then transformed into colour space,and the first-order moment,second-order moment and third-order moment colour features of the three channels are extracted,as well as the texture features of the RVB component with 22 quadratic statistics of the grey covariance matrix to obtain a 31-dimensional feature vector.Secondly,two layers of Haar wavelet decomposition are performed on the super-resolution image,and the grey-scale co-generated matrix texture features are extracted for each of the three high-frequency components of each layer to form a 132-dimensional multi-scale texture feature vector.Finally,the colour features,the grey-scale co-generated matrix texture features and the fused features extracted by the different methods are classified and analysed using a jointly trained classifier.Based on the experimental results,it can be seen that the fused features have a greater classification accuracy for images with different magnifications.The classification accuracy of 40×,100×,200× and 400× images are 94.00%,92.33%,94.21% and 92.14% respectively,indicating that the fusion of low-dimensional features can better classify the benign and malignant breast cancer histopathological images.Moreover,super-resolution can better increase the information such as texture details of the image,which is useful for image extraction of colour moment features as well as GLCM features.