Semantic Segmentation Of Breast Tumors Based On DenseNet

Breast cancer is one of the cancers.The cancers with the most fatal rate which are impairing the health of millions of women globally.If it can be detected in the early-stage cancer and treated,the cure rate of patients will be significantly improved.There are many methods to detect breast cancer,but ultrasound is the most common because it is low cost,noninvasive and portable.Ultrasound breast cancer detection methods are divided into manual,semi-automatic and automatic.The latter two methods both need experienced doctors to designate the region of interest and boundary of the tumor.However,the doctors’ subjectivity will have a negative impact on tumor segmentation results,computer-aided diagnosis（CAD）is proposed to obtain objective results.In recent studies,the deep learning method has been used in automatic detection of ultrasonic breast images,but this method requires a large amount of sample data to converge the model.But it is difficult to obtain sample data marked by doctors.To find out a solution to the above problems,this paper proposes an automatic segmentation method used convolutional neural network.The main research work consists of the following two parts:1.In this study,ultrasound breast image data from the Warsaw Cancer Institute.DenseNet are selected to segmentation.By using dense neural blocks to enhance information flow and feature reuse,the network can directly obtain effective tumor information regions without dividing regions of interest.2.Based on the parameters obtained from the data training of the Warsaw Cancer Institute,the clinical ultrasound data set from Kunming Cancer Hospital was used instead,which was labeled by doctors and enhanced to obtain better semantic segmentation results.Finally,the training accuracy of dense neural network can reach 99.2% while that of traditional neural network is only 85.6%.As for the segmentation result,The DICE parameters of Open Access Series of Breast Ultrasound Data（OASBUD）built on the basis of the Institute of Basic Technology of the Polish Academy of Sciences were up to83.29%,while the DICE parameters of Data from Kunming Cancer Hospital were up to86.03%.The segmentation results were more accurate,and the noise interference was significantly reduced.More importantly,the robustness and generalization ability are greatly improved.