Segmentation Of Breast Ultrasound Image Based On Contour Optimization Algorithm And Deep Learning

Breast cancer,as one of the most common types of cancer in contemporary women,has been a fatal threat to women’s life and health.Early screening of breast can effectively reduce the harm of breast diseases and have been widely implemented in many countries.Among them,ultrasound screening is the first method of breast tumor screening.In the process of breast examination,doctors analyze and judge patients’ breast through the characteristics of ultrasound images.However,the accuracy of diagnosis results depends on doctors’ subjective judgment.The development of computer aided diagnosis system in imaging examination has proved that breast aided diagnosis system can effectively help radiologists to improve the accuracy of breast lesions diagnosis.In breast aided diagnosis system,the accuracy of breast image segmentation has a great impact on the diagnosis results,because breast image segmentation can provide the shape,size,and morphology of the surrounding tissues of the breast lesions,which can help the system to classify the lesions and provide doctors with Bi-RADS grading suggestions.There are many methods for image segmentation,including traditional segmentation method and deep learning segmentation method.Among them,traditional segmentation methods mainly extract and segment specific features according to the characteristics of medical images,while deep learning segmentation methods mainly achieve the purpose of segmentation through the classification of pixels on medical images.In the research process of medical image segmentation,many researchers improve the segmentation accuracy by upgrading watershed,clustering algorithm,active contour model and other algorithms,and some use FCN,U-Net and other deep learning frameworks to improve the accuracy of medical image segmentation.However,breast images are characterized by high noise,blurred edges,and complex lesions,so the above methods still need to be improved in the segmentation of breast ultrasound images.For improving the accuracy of lesions segmentation in breast ultrasound images,this paper mainly designs a segmentation framework combining active contour model and deep learning antagonism mechanism.Meanwhile,a large number of experiments prove the effectiveness of this framework.The segmentation framework proposed in this paper uses conditional adversarial network as the main framework.A new Deformed U-Net and active contour module are used as generators.In this network,we segment breast ultrasound images at pixel level,and optimize tumor lesion edges with active contour model.And each optimization result provides loss feedback for Deformed U-Net,thus improving the classification effect of edge pixels of Deformed U-Net.Markov discriminator is used in the discriminator,and the discriminator results provide loss feedback for the segmentation network,and a more optimized segmentation network is obtained by cross training the discriminator and the segmentation network.The main innovations of the framework are:(1)Deep learning and traditional algorithm are combined to segment breast ultrasound images effectively.(2)Using active contour modules to highlight ultrasound tumor edges,forming a bidirectional optimization with a Deformed U-Net.(3)Antagonism mechanism is adopted to improve the precision of segmentation network.(4)The loss of discriminator,Deformed U-Net and active contour module is used for collaborative optimization of the whole network.The data set for this study is provided by the Department of Radiology,The First Affiliated Hospital of Shantou University.In order to verify the validity of the proposed framework in breast ultrasound image segmentation,three related experiments were designed in this paper:(1)the lesions segmentation in Breast ultrasound images based on the Deformed U-Net;(2)the effect of adding active contour model on breast ultrasound image segmentation;(3)the influence of training with different discriminators on segmentation network.Experiment(1)uses a variety of network frameworks to segment breast ultrasound images.Through qualitative and quantitative analysis of the experimental results,we can find that,comparing with FCN-16 s,FCN-8s,U-Net,U-Net with deformable convolution,and U-Net with ASPP,the results of breast ultrasound image segmentation using Deformed U-Net show obvious advantages.In experiment(2),a comparison experiment was set before and after the active contour model was added.Through qualitative and quantitative analysis,it is found that the segmentation network showed better segmentation performance after the active contour model is added.In experiment(3),different discriminators are used to train the segmentation network.Through qualitative and quantitative analysis of the experimental results show that the addition of Markov discriminator to the training of breast ultrasound image segmentation network can provide certain discrimination loss,which is beneficial to improve the segmentation performance of the network.Compared with other deep learning models,the proposed network based on antagonism and contour optimization can accurately segment the lesions in breast ultrasound images and obtain significantly better segmentation results and segmentation indicators: Dice coefficient,Pixel Accuracy,Precision,Mean-intersection-over-union,Recall,Specificity and F1 score reach 89.7%,98.1%,86.3% 82.2%,94.7%,98.5% and 89.7%.Therefore,we believe that this segmentation framework has good clinical diagnostic prospects.