Breast Cancer Prognosis Prediction Based On Deep Learning

Accurate diagnosis and early prognosis of breast cancer can improve the survival rate of patients.In clinical,the treatment for breast cancer often includes neoadjuvant therapy,which means reducing the size of the tumor and increasing the likelihood of breast-conserving surgery.However,only a proportion of patients showed a pathological complete response to neoadjuvant therapy.Therefore,it will be timeconsuming and high risk for negative patients.It is essential to find an accurate strategy to predict neoadjuvant treatment prognosis.A reliable predictive solution is to use medical imaging techniques such as MRI(Magnetic resonance imaging)to build a CAD(Computer-assisted diagnosis)system.Nevertheless,there are two difficulties,one is to establish an automatic tumor segmentation model and the other is to design efficient image features to predict neoadjuvant treatment response.Therefore,the paper focuses on the following research.(1)The proposed SLAPNet is composed of gaussian pyramid and semantic pyramid.The gaussian pyramid is formed by a gaussian filter that smoothly samples the image layer by layer.The filter operator denoises the image on the one hand and blurs the details on the other,thus highlighting the large structural features of the image.In this paper,the gaussian pyramid was used for creating multi-scale inputs to enable the model to notice not only the global image features such as shape and gray-level distribution but also the local image features such as edges and textures.By combining multi-level features,SLAPNet is more robust and versatile,and more powerful in handling irregular object segmentation.The semantic pyramid first extracts deep semantic features from multi-level inputs by U-Net and then connects adjacent layers to pass deep semantic features between different layers.The strategy incorporates multi-semantic layer features and multi-level features to improve the performance of the model.The results showed that the pyramid model performed best in multicenter breast cancer tumor segmentation.(2)Unsupervised transfer learning features were proposed to predict the response to neoadjuvant therapy in breast cancer patients.The image feature maps were first learned in an unsupervised manner by Convolutional Restricted Boltzmann Machine(CRBM)model,and then the feature maps were input into a pre-trained VGG19 net to further extract semantic features.Afterward,least absolute shrinkage and selection operator(LASSO)regression was implemented for feature dimensionality reduction.Finally,support vector machines(SVM)and logistic regression(LR)were trained with selected features to predict neoadjuvant treatment response.The results suggested that the prediction with unsupervised and transfer learning-based image features before neoadjuvant treatment for breast cancer was helpful to accurately predict the response to the therapy,and it helped guide the personalized treatment of breast cancer.