Research On Automated Segmentation For 3D Breast Ultrasound Image

Breast cancer is the most common diagnosed cancers and is one of the leading causes of cancer mortality in women.Ultrasound imaging is one of the chief methods for early diagnosis of breast tumors due to its noninvasive,real-time visualization,low cost and easy operation.However,the analysis of ultrasound images is often limited by the clinicians' theoretical basis and experience,relying solely on manual interpretation would lead to diagnosis with subjectivity and deviation.Segmentation of breast ultrasound images can provide more accurate and objective opinions for clinicians,which is an important part of computer-aided diagnostic system.Ultrasound imaging has some inherent defects such as speckle noise,low contrast and low resolution,which increase the difficulty of image segmentation.This paper devotes to research a CNN-based segmentation method,which can segment the 3D breast ultrasound images in high precision.Based on the convolutional neural network algorithm in deep learning field,this paper proposes an automatic segmentation method of breast ultrasound images,which transforms the segmentation problem into the classification of each pixel in the image.In the proposed method,we first form the training set from the clinician's mark,select the image patch centered on a certain pixel as the input of the convolutional neural network,and the actual class of the pixel as the label.Then we use convolutional neural network as a pixel classifier and train the network,the final test set accuracy is more than 80%after training.When the trained model is used to segment the image,the class of each pixel can be calculated from each image patch,so that the final segmentation result is obtained.The method proposed in this paper can correctly distinguish skin,glandular tissues and tumors in breast ultrasound images,the shape and edge of the segmentation results is similar to the standard results marked by clinicians,and also achieved good results in the quantitative evaluation.In the experiment,the optimal model is selected by comparing the results of the different network parameters.In addition,the proposed method can get more accurate segmentation results compared to some other methods.In this paper,convolutional neural network is used to deal with the complex ultrasound images.Through a large training set,the computer learned the complex relationship between the images and the results,which is more similar to the thinking mode of human brain.The segmentation process does not need human intervention and prior knowledge,can be fully automated,and can get better versatility when having sufficient training data.The proposed method provides a new prospect for the computer-aided diagnosis system.