Research On Automated Breast Cancer Histopathological Grading Through Digital Mammograms Based On CNN

The accurate histopathological grade of breast cancer reflects the internal characteristics of the tumor.It is an important prognostic indicator of malignant tumor,which is closely related to the recurrence interval of cancer and the length of the survival period.In clinic,the histopathological grade of breast cancer is based on the pathological images obtained from breast biopsy.Digital mammography screening is the most convenient and noninvasive examination with high resolution.Therefore,directly predicting the histopathological grade of breast cancer from digital mammograms has great clinical significance.The representation of the tumor region is very similar and has little difference and the tumor area in the molybdenum target mammogram is relatively small.In order to reduce the influence of the non-tumor region on the classification algorithm,the discrimination of histopathological grade is implemented based on the segmented tumor.Convolutional Neural Network(CNN)has powerful feature extraction and expression ability.It can automatically extract high level semantic features from the image,which is helpful for the segmentation of breast cancer and the classification of histopathological grade.In the digital mammograms,malignant tumors have the characteristics of diverse shapes and different sizes.Because digital mammogram data with accurate pathological grades are not easy to be collected and the tumor outlines need to be labeled by professional Radiology doctors,the labeled data is scarce,which makes the training of the CNN based breast cancer segmentation and pathological grading algorithm difficult.In view of the above problems,this research considers to improve the accuracy of breast cancer segmentation and histopathological grade classification algorithm based on the fusion of multi-scale and multilevel feature.The main achievements of the paper are as follows:1.Aiming at the various shape and size of malignant tumor and the relatively small tumor area in the mammograms,fully convolutional network based on Atrous Spatial Pyramid Pooling(ASPP)is proposed by considering the fusion of multi-scale image information to improve the precision of breast cancer automatic segmentation.Firstly,reduce the number of the total downsampling layers to improve the recognition ability of the small size tumor.Secondly,the ASPP module which combines the different receptive field information of the network,that is,the multiscale information of the image,is added to the whole network to improve the network segmentation ability for different sizes of tumors.Finally,the proportion of tumor pixels in the image is calculated and added to loss function to enhance the tumor loss.Through qualitative and quantitative experimental results,it is proved that the algorithm proposed in this chapter has high segmentation accuracy for different sizes and shapes of digital mammograms,and provides accurate tumor areas for subsequent histopathological grade classification.2.Due to the scarce annotated mammograms data,CNN training is easy to be overfitted,so this study proposes automated histopathological grading of breast cancer on digital mammograms based on S-DenseNet.Firstly,the train labeling dataset is enlarged by data augmentation.Secondly,through analyzing the main growing way of the training parameters of DenseNet which connects all layers directly with each other,a light and deep network with less training parameters is modified to prevent overfitting,and at the same time maintain the layer connection and feature reuse manner.The quantitative analysis shows that S-DenseNet network has higher accuracy than the traditional classifier and other CNN networks in the classification task of breast cancer histopathological grade.3.Considering the difficulty of distinguishing different histopathological grades on digital mammograms and the lack of low-level details of tumor extracting by CNN,multi-level features combined end-to-end learning algorithm for automated histopathological grading of breast cancer on digital mammograms is proposed in order to further improve the accuracy of histopathological grading.Firstly,a variety of low-level detail features are extracted based on the segmented tumor,and then a supervised LASSO logistic regression is proposed to select the features that are useful for the classification.Finally,based on the pre-trained S-DenseNet network model,the selected low-level features are incorporated into the network for combined training.The high-level semantic features,the low-level features and histopathological grade are fused to update the network parameters,making the CNN feature extraction part of the fusion network more focus on extracting the high-level semantic features which complement the low-level features.Quantitative comparison experiments show that the algorithm has better classification effect in breast cancer histopathological grading tasks.