Breast Ultrasound Image Tumor Segmentation Based On UNet

Due to the different imaging principles and imaging equipment of ultrasonic images,there are differences in contrast and resolution of ultrasonic images obtained,and the different shapes of breast tumor lesions of different individuals,which lead to the low accuracy of traditional ultrasonic image segmentation algorithm.In this context,deep learning algorithm is adopted in this paper to segment breast ultrasound images,so as to improve the accuracy of segmentation of breast tumor lesion areas.Specific work contents are as follows.Firstly,the CSA(Channel and Spatial Attention)module is proposed and the loss function is improved.Since the spatial attention module can obtain the spatial information in the feature map,and the channel attention module can obtain the channel domain information in the image feature,the method of combining the two is adopted,so that both the channel and the spatial features of the image can be preserved at the same time.The Binary Cross Entropy Loss(Bce Loss)and Dice Loss were combined to improve the loss function.Secondly,a module with intensive cavity convolution is proposed.Continuous downsampling in the CSA-Unet(Channel and Spatial Attention Unet)network structure reduces the resolution of ultrasonic images and leads to the loss of image information.In order to alleviate this problem,intensive void convolution was incorporated into CSA-Unet to increase the model’s sensitivity field and reduce the loss of semantic information in the process of decreasing ultrasonic image resolution without changing ultrasonic image resolution.Finally,the structure of the dual encoder is integrated into the model.Because the deep learning method will inevitably lose the edge information of breast tumor features during subsampling,and the edge of ultrasonic image itself is not clear,resulting in limited segmentation performance of the model.Using the dual encoder structure,the edge details extracted by the left encoder are supplemented by the right encoder,so that the whole model can obtain more edge features.