Early Diagnosis For Breast Cancer With Deep Learning

In the worldwide,breast cancer with its high incidence,high mortality has been a serious threat to women's health.And the incidence of breast cancer in recent years has been on the rise.Because of its uncertain pathogenesis and disease concealment,making early breast cancer is difficult to be found.Dynamic contrast-enhanced magnetic resonance imaging(DCE-MRI)has been widely used in the diagnosis of early-stage breast cancer.Deep learning based on biological visual principles can automatically learn the hierarchical characteristics of data,making it of great success in the image,audio,natural language processing and other aspects.Recently,people have been exploring the application of deep learning in the field of biomedicine,and have achieved some results.The research of this thesis is mainly based on the application of the deep learning method in the diagnosis of early breast cancer,and explores the early diagnosis of breast cancer using imaging data of different modalities based on different deep learning methods.The main contents of this paper are as follows:(1)Based on unsupervised learning stack auto-encoding feature extraction methods and classification.Unsupervised Learning Stacked Autoencoder(SAE)was used to extract features and classify.First of all,preprocess of experimental data,including extraction of ROI and PCA whitening.Then by the use of unsupervised layer-by-layer training to extract different hierarchical of features,and finally using Softmax classifier for early benign and malignant breast cancer classification.(2)Early diagnosis of breast cancer based on three-dimensional convolution neural network.The three-dimensional convolutional neural network is constructed and the dataset are expanded by translation,rotation and mirror image.And then using two-dimensional and three-dimensional CNNs for classification and prediction.Three-dimensional enhanced image sequence and enhancement rate images were used to identify and classify early breast cancer.(3)research on transfer learning to discriminate benign and malignant breast cancer.To study the application of transfer learning model in benign and malignant classification of early-stage breast cancer,the pre-trained convolution neural network on the large data set(such as ImageNet)was used as the feature extractor of the bottom and middle layers and then transfer to the MRI image dataset by fine-tuning the model parameters for classification.From the results of this research,the method proposed in this paper can better identify the classification labels of early stage breast cancer.The AUC of SDAE for early diagnosis of breast cancer was 0.85.The sensitivity,specificity and AUC of based on 3DCNN classification for early breast cancer were 0.82,0.74 and 0.80,respectively.In migration network model classification experiments,AUC,sensitivity and specificity were 0.86,0.85 and 0.81,respectively.In the future,if the deep learning method automatically learns the different hierarchical of features fussed into the MRI assisted diagnosis system,will help improve the system performance,which has good application prospects.