Convolutional Networks Based Classification And Segmentation Of Breast Ultrasound Lesions

Breast cancer is the second leading cause of death among all female diseases around the world and causes massive deaths every year.Breast ultrasound diagnosis has becoming a common means to diagnose breast tumors nowadays,and how to identity a breast tumor as benign or malignant and segment its focal area accurately in ultrasound images are the cores of computer-aided breast ultrasound diagnosis systems.In this thesis,due to the low signal-noise ratio and low contrast between breast tumors and surrounding normal tissues,constructed defects,blurred edges and even loss of edge information in breast ultrasound images,data-driven based convolutional networks are applied for extracting feature representations of breast lesions,then the classification for identifying breast tumors as benign or malignant and segmentation of breast tumors' focal areas are handled by these features.The network for classification problem,called the RecNet,belongs to residual networks.In the RecNet,residual connections are utilized to force existing features to fuse other remaining information in the forward pass,and error terms of neurons in deep layers can be transmitted effectively to neurons in shallow layers in the backward pass,which can alleviate the occurrence of gradient vanishing.The network for segmentation problem,called the SegNet,combines the characteristics of fully convolutional network(FCN)framework and high-resolution network(HRN)framework.In the SegNet,the encoding subnetwork is used to obtain high-level semantic features of input images,and the fine-grained information is restored through the decoding subnetwork.At the meantime,dilated convolution operators are utilized to expanding the local receptive fields of neurons in deep layers,which is helpful to capture the semantic information of breast lesions in input images.On the breast ultrasound dataset collected by Nanjing Drum Tower Hospital,the RecNet achieves an accuracy of 84.0%,and a dice score of 93.1% is obtained by the SegNet.Experiments show that the RecNet and the SegNet can recognize and segment breast lesions automatically and accurately,thus reducing the workload of physicians.