Breast Ultrasound Image Segmentation Based On Semi Supervised Learning

Breast cancer seriously threatens the health and safety of women worldwide and is the leading cause of death for women worldwide.Early breast cancer screening plays a crucialrole in its treatment.Ultrasonography has become one of the main methods of breast cancer diagnosis due to its advantages of safety,low price,no radiation,and easy operation.Computer-aided diagnosis technology has significantly improved doctors’ diagnostic efficiency and accuracy,especially with the rapid development of data-driven deep learning technology.Lesion segmentation is an important step in computer-aided diagnosis.However,issues such as the high cost of labeling ultrasound images and medical ethics privacy limit the size of the dataset.In addition,the problems of low contrast,high speckle noise,and blurred boundaries of ultrasound images also affect the performance of the learned model.Starting from the realistic problem of insufficient labeled data,this thesis combines generative adversarial learning and limited labeled data to design a suitable semi-supervised method to achieve accurate breast ultrasound tumor segmentation.First,this thesis proposes a semi-supervised breast ultrasound image segmentation network based on adversarial learning to perform better with a small amount of labeled data.The network consists of two generators and a discriminator,and the two generators and the discriminator are trained against each other.Model training is divided into two stages.In the first stage,the model is only loaded with labeled images,and the generator and discriminator are initialized.In the second stage,unlabeled images are added,the discriminator screens the predicted segmentation masks of the unlabeled images,and the filtered segmentation masks are used as supervision signals,so that the two generators form a mutually supervised training mode.Only a small amount of labeled data is used to make the model achieve segmentation performance close to fully supervised training.Secondly,in order to further improve the accuracy of breast ultrasound segmentation,this thesis explores the internal relationship between breast ultrasound classification and segmentation tasks and proposes a joint learning framework for breast ultrasound segmentation and classification.The framework includes feature sharing paths,segmentation sub-networks and classification sub-networks.The segmentation and classification networks share the feature extraction path,which can reduce the network parameters multiplied by dual tasks.The segmentation sub-network uses the obtained prediction mask to guide the training of the classification sub-network with its position,boundary,and shape information.The classification sub-network obtains the deep and shallow class activation maps and corrects the segmentation sub-network.The mutual promotion of segmentation and classification tasks is achieved,the classification accuracy is increased by 3.11%,and the intersection ratio of segmentation is increased by nearly 2%.Finally,this thesis develops a breast ultrasound intelligent auxiliary diagnosis system based on the above algorithms,doctors’ needs,and diagnostic procedures.Its main functions include automatic segmentation of lesions by hand,automatic diagnosis of lesions,and comparison of similar cases.The system can reduce the workload of doctors,provide doctors with more reliable diagnosis references,and improve the reliability and efficiency of diagnoses.