Deep Convolutional Networks For Automated Tumor Segmentation Of Breast DCE-MRI

As a disease with high morbidity and mortality among women,breast cancer seriously threatens the health of many women.In many applications of DCE-MRI,quantitative description of two-dimensional(2D)diameter and three-dimensional(3D)volume of breast tumors is of great clinical value,such as neoadjuvant chemotherapy can guides follow-up treatment by judging changes in tumor focus areas at different stages.Automated segmentation of tumors is also a key component of quantitative evaluation of the degree of breast malignancies,such as quantitative analysis of 2D or 3D morphology of tumors,texture of tumors,etc.by using imaging,which depends heavily on two-dimensional or three-dimensional segmentation of tumors.However,the precise segmentation of tumor regions obtained by most influential studies still relies on manual marking,which is time-consuming and laborious for doctors.Therefore,the research of automatic segmentation algorithm plays an important role in assisting doctors in diagnosis and imaging research.At present,radiomics mainly uses the maximum cross-section of breast tumors labeled by doctors.A method based on deep convolution neural network is proposed to realize the 2D segmentation of the maximum cross-section of breast tumors towards this goal.This method first roughly detects the approximate regions of breast tumors through the deep detection network,and then extracts these regions into the deep segmentation network for precise segmentation.The average Dice coefficient,average specificity and average sensitivity of this method in testing sets which contains 160 cases are 0.81,0.78 and 0.86.In order to further verify that the automatic segmentation results can provide help for the follow-up imaging research,some related experiments were carried out.Experiments show that the results obtained by using the automatically segmented regions are basically consistent with those obtained by using manually labeled regions.This demonstrates that the automatic segmentation method can provide reliable 2D breast segmentation data in molecular analysis tasks of breast tumors.Although the 2D segmentation method proposed in this paper can achieve better results in the maximum cross-section,the 2D tumor region still can not reflect the shape of the tumor region more comprehensively.Therefore,based on dilated convolution module,a 3D convolution neural network A-Net is proposed to realize 3D segmentation of breast tumors.The average Dice coefficient,average VOE value,average ASSD value and average aRVD value of this method are 0.82,0.11,0.85 and 0.10 respectively in testing set which contains 22 cases.These indexes show that this method has great advantages over 3D U-Net and V-Net.2D and segmentation algorithms of breast tumors studied in this paper can provide effective solutions for breast imaging research,artificial intelligence-assisted diagnosis and treatment of breast tumors and the establishment of prognostic system.