The Research On Segmentation Of Breast Ultrasound Image Based On Superpixel Classification And Multiscale Attention Mechanism

In recent years,breast cancer has become the leading cause of cancer in women and it poses a great threat to women’s lives and health.As a non-invasive,fast and cheap technology,ultrasound imaging has been regarded as an important method for the diagnosis of breast cancer.In order to reduce the subjectivity of diagnosis and reduce the workload of doctors,computer-aided diagnosis systems are developed rapidly.The results of image segmentation play a key role in the final judgment of the systems.However,due to the high noise,low contrast and poor uniformity of the ultrasound images,the segmentation of breast ultrasound images is still a challenging task.At present,the segmentation methods based on traditional machine learning mostly implement the classification by extracting low-level features of small patches of images.However,the segmented edges obtained by these methods are often discontinuous,which is difficult to describe the morphology of breast tumors accurately.In order to achieve more accurate segmentation,two segmentation methods of breast ultrasound images are proposed in this paper.Considering that high-level semantic features are helpful for the location and judgment of breast tumors,Kmeans is utilized to construct a bag-of-words model based on the low-level image features of superpixels to obtain higher-level semantic features.After the semantic classification,a reclassification method is designed to correct some misclassified superpixels and further improves the segmentation performance.Then,in order to further reduce the semantic gap between segmentation targets and features,a deep learning model based on a multi-scale attention mechanism is presented,which combines multi-scale features and attention mechanism to the fully convolutional neural network.In the proposed model,the idea of the residual learning unit is adopted and a multi-scale residual module is designed,which uses the convolution kernels of different sizes to construct multiple shortcuts to enhance the feature extraction capability of the model.In addition,an upsampling module based on the multi-scale attention mechanism is also proposed.The low-level features are selected through the attention mechanism and then added to the high-level features to enhance the restoration of the image spatial information.320 clinical breast ultrasound images are collected as the experimental dataset and five-fold cross-validation is performed on the two methods proposed in this paper.Theexperiment results show that the proposed method based on superpixel classification has better comprehensive segmentation performance than four traditional image segmentation methods and is close to the fully convolutional neural networks.The proposed method based on multi-scale attention mechanism is superior to FCN,Seg Net and U-Net and achieves the largest F1-score value.Both methods proposed in this paper can accurately extract the tumor edges and the edges are continuous and smooth,which are close to the contours labelled by the doctors.