DCE-MRI Breast Cancer Diagnosis And Molecular Subtypes Prediction Based On Deep Learning

With the development of economy and the improvement of living standard,people are attaching more and more importance onto health,especially the accurate and fast diagnosis and personalized treatment of diseases.Breast cancer is the most serious malignant tumor in women's health.Dynamic contrast-enhanced magnetic resonance imaging(DCE-MRI)has been widely used in the early detection and diagnosis of breast cancer.In breast cancer diagnosis,computer aided diagnosis(CAD)system based on breast DCE-MRI images is commonly developed by extracting a set of handcrafted features from lesion area.While,estimating these features requires rich experience and has the subjectivity.During breast cancer treatment,doctors make suitable treatment of patients with the personalized according to the different tumor subtypes.The breast cancer molecular subtypes are confirmed by immunohistochemical results of breast cancer tissues.The performance process is complicated and traumatic injury.Because of the limites,the above breast cancer diagnosis and treatment methods can not satisfy the accurate diagnosis and personalized treatment.Therefore,the breast cancer diagnosis and treatment need to further improvement.At present,deep learning can automatically learns features in different depths from a given image.This method has been widely applied in the field of pattern recognition,which has achieved better effects than the traditional method with handcrafted features.Based on the existing methods and experiences of deep learning,this paper explored the application of convolution neural network in the breast cancer diagnosis and molecular types prediction with DEC-MRI.This paper is mainly as follows:First,we extracted the object region that contains the lesion area from original DCE-MRI image based on the lesion information marked by the physician.The expanded data were then established by zooming,rotating,sliding window and mirroring which may included as much lesion features as possible for deep learning.Secondly,this paper constructed a two-dimensional convolution neural network model and a three-dimensional convolution neural network model respectively to analyze lesions in different dimensions for prediction of breast tumors.The results showed that the two-dimensional convolutional neural network with more convolutional layers achieves better classification and prediction performance than 3D model.When the model parameters were generated by random initialization method,performance of 3D convolutional neural network is better than 2D model for prediction of benign or malignant tumor.When using migration initialization method to produce the model parameters,the performance of two-dimensional convolutional neural network in prediction of breast tumors improved greatly.Finally,we combined the above two-dimensional convolutional neural network and three-dimensional convolutional neural network respectively with the enhanced information of time dimension to predict the breast tumors molecular subtype.Results shown that performance of 2D model with migration initialization method was better than other methods.Results also shown that performance of classification for benign or malignant breast tumor is better than prediction of breast tumor molecular subtype with deep learning architecture.This paper explored breast cancer diagnosis and molecular typing predictions based on deep learning method with DCE-MRI.Results indicated that deep learning method has beneficial effect on classification of benign or malignant of breast lesions,and has definite effect on prediction of breast tumor molecular subtype.Of particular importance is the potential value of prediction for breast tumor molecular subtype with non-invasive method.