Research On Multi-objective Semantic Segmentation Of Ultrasonic Images

Breast tumor is a common disease which threatens the health of women in recent years.Clinical studies has been shown that most of the early breast tumors can be successfully cured if they can be found and treated in time.Therefore,the medical diagnosis of breast tumors is particularly important.Ultrasound imaging technology uses ultrasound beams to pass through the human body and forms ultrasound images based on the strength of acoustic reflections.It is convenient and intuitive to observe diseases such as tumors and deformities in the human body.Today,the technology has been widely used in the detection of breast tumors with its advantages of non-invasive,painless,and low cost.In order to reduce the misdiagnosis rate of breast ultrasound images and assist doctors in the work,the computer can automatically identify the tumor area and surrounding tissue structure is the main purpose of this study.In the field of medical image processing,the main way to identify objects is to get the target area by image segmentation.The method in this paper is to make semantic segmentation of images.This segmentation method can not only determine the location of the target,but also identify the categories of the target.The classification model is used to calculate the classification probability distribution of each pixel,and the full connection condition is used to optimize the whole image to determine the category of each pixel and achieve the purpose of semantic segmentation.This paper first summarizes the research status of image segmentation at home and abroad,and then proposes the idea of the image semantic segmentation method in this paper.It extracts a variety of different features from the obtained breast ultrasound image samples,and they were respectively trained using the support vector machine model,and their image segmentation effect on the test sample was compared.The deep features extracted by the full convolutional neural network in the experiment had better segmentation accuracy.In this paper,the pixel class probability distribution of image segmentation is interpolated and enlarged to get the coarse segmentation of the image.Then a fully connected conditional random field model based on medical priori constraints is proposed.This model takes into account the hierarchical features and adjacent relationships of tissue and organs in breast ultrasound images,and increases the distance constraints between categories.It can punish the segmentation results which are not consistent with medical knowledge,so as to correct the error segmentation caused by the classification model in rough segmentation.Experiments show that the method in this paper candeal with the problems of unclear image features,SVM classifiers that have a violation of medical common sense,and can achieve better segmentation results on breast ultrasound images.