Research On Breast Cancer Histopathological Images Recognition Based On Feature Fusion

Breast cancer is the biggest threat to women's health.It is considered to be the first killer of women today.Early diagnosis of breast cancer is the key to improving the cure rate and survival rate of patients.The results of pathological examination can provide the final diagnosis basis,which is considered as the gold standard for cancer diagnosis.In recent years,the rapid development of computer-aided diagnosis technology has facilitated the diagnosis of breast cancer,with high diagnostic efficiency and more objective and scientific diagnostic results.However,the current computer-aided diagnosis system is mainly based on the computer-aided diagnosis technology of imaging,while the computer-aided diagnosis technology of digital pathology is still in the stage of laboratory research.It is an important research topic to study the automatic recognition of breast cancer histopathological images and to promote the rapid development of digital pathology computer-aided diagnosis technology.In this dissertation,the breast cancer histopathological images recognition of benign and malignant recognition and sub-classes recognition problems under the conditions of magnification specific and magnification independent were discussed based on the Brea KHis dataset,the details are as follows:(1)In view of the insufficient utilization of breast cancer histopathological image features and the large amount of calculation of high dimensional features,a three-channel low dimensional features based breast cancer histopathological images recognition method was proposed for benign and malignant recognition.First,the original images were separated into three channels of R,G,and B.And the three channel features of gray level co-occurrence matrix on one direction,gray level co-occurrence matrix on four directions,average pixel value of each channel,Hu invariant moment,wavelet,Tamura,local binary pattern,completed local binary pattern,Gabor and histogram of oriented gradient were extracted,respectively.Then,the features of the three channels were fused,and the benign and malignant recognition of breast cancer histopathological images was realized through the support vector machine.Experimental results show that the discriminability of three-channel features is better than gray-level features.R channel features have a greater impact on the recognition results with magnifications of 40×,100×,and 200×,while for the images at 400×,it is more dependent on the B channel features.In addition,the comparative experimental results show that low dimensional features can achieve high recognition accuracy while reducing running time.(2)Aiming at the problems that the design process of handcrafted features is complicated and requires a large amount of prior knowledge,a deep learning based breast cancer histopathological images recognition method was proposed for benign and malignant recognition that are magnification specific and magnification independent.To highlight the importance of different channel features,the SE-Res Net module was introduced into the traditional convolutional neural network to obtain more detailed information of the target through the attention mechanism.Increasing the feature representation while assigning different weights to different channel features to highlight the effective information and suppress the useless information.Different from traditional convolutional neural networks for image processing,this model used a 3×1 convolution kernel instead of a square convolution kernel to reduce the parameters and accelerate the training speed of the model.To make full use of the features of different layers of the network,the feature maps of different sizes were upsampled to 224×224 through bilinear interpolation,and the features of different channels were fused through global average pooling to further improve the effectiveness of the features and achieve more accurate recognition of breast cancer histopathological images.(3)Considering the low recognition accuracy of the traditional convolutional neural networks training from scratch,and a large number of labeled training samples are required,while the labeling of histopathological images is a difficult task,a method based on pre-trained convolutional neural networks was proposed for breast cancer histopathological images recognition,which was used for benign and malignant recognition and sub-classes recognition under the conditions of magnification specific and magnification independent.On the one hand,to preserve the spatial information of the images as much as possible,the deep convolutional layer features were extracted as deep semantic features through the pre-trained Dense Net201 network.On the other hand,the gray level co-occurrence matrix features of the three channels were extracted from the original images.Then the deep semantic features and texture features were fused to realize the breast cancer histopathological images recognition through support vector machine.Compared with the 7 baseline pre-trained models of Alex Net,VGG16,Res Net50,Goog Le Net,Dense Net201,Squeeze Net and Inception-Res Net-v2,the proposed method obtained higher image-level accuracy and patient-level accuracy.Finally,the factors affecting the recognition of breast cancer histopathological images were analyzed based on Brea KHis dataset.By comparing the accuracy with other literatures,the effectiveness of the three methods proposed in this dissertation for recognition of breast cancer histopathological images is fully verified.