Segmentation And Diagnosis Of Tumor From Breast MRI Based On Deep Neural Networks

Breast cancer is one of the most common breast cancers in women today,and its incidence is increasing year by year.Early detection of breast cancer and reasonable treatment methods are of great significance for improving the survival rate of patients with the disease.Due to its high sensitivity and specificity,Magnetic Resonance Imaging(MRI)technology has become an important method for the diagnosis of breast cancer and been widely used in clinical practice.Tumor segmentation and diagnosis in breast MRI have become important research content in the field of Computer-Aided Diagnosis(CAD).In this thesis,according to the characteristics of breast MRI image,a tumor segmentation algorithm and two tumor diagnosis algorithms based on deep neural networks are proposed.The main research contents include:(1)A breast tumor segmentation algorithm based on mixed 2D and 3D convolution network with multi-scale context is proposed.Based on U-Net,a mixed 2D and 3D convolution module is added to extract the intra-context of MRI slices,and atrous convolutions with different sampling rates are added at the end of the network encoding component to extract multi-scale image features.At the same time,a multi-channel image representation is proposed to fuse the time series information in the Dynamic contrast-enhanced MRI(DCE-MRI)into a single image as the input of the network to obtain a more accurate segmentation result.Experimental results show that the Dice Similarity Coefficient(DSC),Sensitivity,and Positive Predictive Value(PPV)of this algorithm are 76.5%,75.9%,and 82.4%,respectively,which outperformed other advanced algorithms in this field.(2)A breast tumor diagnosis algorithm based on sequence division and the multi-model integrated network is proposed.By referring to the pathological signs that radiologists pay attention to in clinical diagnosis,the DCE-MRI images are divided to extract more comprehensive and effective pathological features.At the same time,in order to obtain the intensity change information of the lesion area before and after the contrast agent intake,a network combining a convolutional neural network and a recurrent neural network is used to extract the spatial and temporal features.Experiments show that the accuracy,sensitivity,and specificity of the algorithm are 83.3%,88.8%,and 71.4%,respectively,which surpasses other advanced algorithms in this field.(3)A breast tumor diagnosis algorithm based on an adaptive weighted multiple sequence integrated network in multi-sequence MRI is proposed.Through the intermediate output of the adaptive weighting module,the weights of different subsequences in diagnosis can be obtained,thereby assisting the radiologist’s diagnosis in a targeted manner.The final experiments show that the algorithm improves the classification accuracy rate by 4.4% compared with the algorithm based on the single-sequence classification.In summary,our algorithms can achieve the tumor segmentation and benign and malignant classification from breast MRI and improve the performance of breast cancer diagnosis.It can also provide good auxiliary diagnostic suggestions for the clinical diagnosis of radiologists.