Breast Abnormality Classification And Lesion Segmentation Based On MRI Images

As one of the cancer diseases that occur more frequently in women,the incidence of breast cancer is increasing year by year and the incidence of the population is gradually getting younger.Early detection and intervention are of great significance for improving patient survival.With its high sensitivity and specificity,Magnetic Resonance Imaging(MRI)technology has become an important method for diagnosing breast cancer and has been widely used in clinical practice.Mammary magnetic resonance imaging mass segmentation and image abnormality diagnosis have also become important research contents in the field of Computer Aided Diagnosi.With the development of science and technology,the diagnostic assistance system based on deep learning has made a big breakthrough in traditional methods,but because MRI breast image has the characteristics of early lesion tissue and normal tissue,it is not easy to distinguish and the edge of the lesion is intertwined with normal tissue,and it is not easy to completely segment,resulting in some problems in the breast assist system.Aiming at the above problems,according to the characteristics of breast MRI imaging,this paper proposes a tumor segmentation algorithm and a diagnostic algorithm for breast image abnormalities based on convolutional neural network,and verifies the effectiveness of the algorithm through experiments,the main research contents include:(1)A shallow classification algorithm D-VGG8 based on atrous convolution is proposed.There is less lesion information in breast MRI images,and the deep network is easy to extract more background features.This imbalance in the proportion between classes will affect the results and cause overfitting.Inspired by the fact that small datasets are prone to overfitting in deep network training,this paper proposes a classification algorithm D-VGG8,which uses the 8-layer VGG algorithm as the basic algorithm,and integrates the convolution operation with atrous convolution to reduce the network level while ensuring to increase the field of view of the network,the algorithm also adds regularization and dropout layers to the convolution process to reduce the occurrence of overfitting.In this paper,the algorithm is compared with Res Net,Dense Net and VGG algorithms with different structures.The classification results of Precision,Recall,F1,and Acc are 0.86,0.85,0.85,and 0.84,respectively,which are 0.04,0.07,and 0.07 higher than VGG16,respectively.0.04,0.04,there is a certain improvement compared with other algorithms.(2)A breast segmentation algorithm SE-UNet++ based on feature compression is proposed.In order to solve the problem that the edge of breast lesions is similar to normal tissue and difficult to complete segmentation,inspired by the attention mechanism,this paper proposes the SE-UNet++ algorithm,which uses the U-Net++algorithm as the basic algorithm,and adds a feature compression excitation module to the network.The algorithm refines features through the attention module,and uses skip connections to connect and reuse features at different levels,which not only solves the semantic gap problem of the U-Net network,but also provides good features for generating predicted images in the upsampling stage.information.In this paper,the algorithm is compared with several U-Net series improved algorithms.The segmentation results are 0.741,0.84,and 0.775 in MIOU,MPA,and SE,respectively,which are 0.15,0.004,and 0.25 higher than the U-Net algorithm.The algorithm has also been improved to some extent.To sum up,the algorithm proposed in this paper can realize the task of segmentation and image classification of breast MRI image lesions,improve the diagnostic performance of key technologies of breast imaging diagnosis assistance system,and also provide better auxiliary diagnosis suggestions for doctors’ diagnosis work.