Study On Segmentation Of Breast Tumor Based On Deep Learning

As a kind of high incidence disease,breast cancer has greatly threatened the global females both in physical and psychological health.The most effective way to treat the disease is the early screening for breast cancer.Mammography is considered as the one of the most common image modalities.However,the various characteristics of the tumor and radiologist’s expertise make the manual annotation complicated and low efficient.The mass is one of the most import representations in breast cancer.The traditional segmented algorithm makes it difficult to extract the mass from the contrast-similar tissue based on the features such as grayscale,morphology,texture.With the widely application of computer-aided diagnosis(CAD)in the medicine,the automated segmentation of breast tumor can provide the homogeneous result for the physician and improve the diagnosis accuracy.This thesis focuses on two algorithms for the segmentation of malignant tumor in mammography with the knowledge of deep learning and image process.The first method proposes a two-stage convolutional neural network based on super-pixel which transforms the segmentation task into the binary classification and realize the tumor segmentation from the super-pixel level to the pixel level.Based on the deep supervision blocks,the second model adapt the architecture of encoder-decoder network which learns from the mammographic image without preprocessing and produce the segmented mask of the same size.The segmented contour is fined and the redundant information is reduced greatly compared with the CNN using image patches.The experiment data is established on the public data-set from the University of South Florida and the private data-set from the radiology department of Zhongnan hospital in Wuhan university.The experiment result shows that both of the two model get great performance on the segmentation of mass on the mammography and achieve competitive segmentation indexes.The research work in this thesis is perspective for the early diagnosis of breast tumor and the later targeted therapy.