Research On Automatic Segmentation Technology Of Breast Tumor Based On Ultrasound Image

The automatic segmentation of breast ultrasound images is an important part of the computer-aided diagnosis system based on breast ultrasound images,which has become a research hotspot in the field of human-computer interaction,widely used in breast cancer diagnosis.Breast ultrasound image segmentation technology,which involves medicine,biology,statistics,computer science and other disciplines,is a very novel and valuable research direction.The quality of ultrasound images is poor and the number is small,and individual breast tumors vary greatly,which greatly increases the difficulty of breast ultrasound image segmentation.In this thesis,we aim to build and train a neural network model suitable for breast ultrasound images to improve the accuracy and practicability of the segmentation.The work content is as follows:(1)In order to improve the importance of breast ultrasound image lesion in the segmentation task,Attention U-Net was improved,and an Attention-enhancing U-Net(AE U-Net)is proposed.First,the network loss function was improved.Based on the output value of the traditional network end,output weights of all attention gate were integrated.Compared with the corresponding standard lesion template,the loss function was calculated to enhance the weight of the lesion area and obtain an accurate network loss value.Secondly,the network training method was improved,and the strategy of combining roughness and fineness was adopted.The overall loss function was used to train the overall network to make the network basically stable,then the partial loss function was used to alternately train the backbone network and the attention gate module in turn.Fine-tuning was used to further improve the accuracy of network parameters.The combination of the two greatly improves the accuracy of segmentation of the breast ultrasound lesion area.Experimental results on a custom breast ultrasound database show that the proposed AE U-Net has an M-IOU of81.24%,Acc of 93.85%,AUC of 93.44%,Precision of 85.88% and Recall of 79.48%.(2)The multiple down-samplings in the AE U-Net greatly reduce the resolution of the feature maps,resulting in the loss of internal data structure and spatial hierarchical information.In order to alleviate the adverse impact of this problem on the image segmentation task,AE U-Net was improved,and an AE U-Net with HDC is proposed.Three sets of HDC with expansion rates of [1,2,5] were incorporated into AE U-Net to replace the fourth down-sampling of the left contraction path in the network structure and the subsequent convolution operation,to replace the first up-sampling of the expansion path on the right and the subsequent convolution operation.This ensures the receptive field of the network while keeping the resolution of the feature map unchanged,reduces the loss of spatial information caused by the reduced resolution of the feature map,improves the accuracy of segmentation results in pixel-level image segmentation tasks.Experimental results on a custom breast ultrasound database show that the AE U-Net with HDC model proposed in this thesis has an M-IOU of 83.08%,Acc of 94.63%,F1 of 83.37%,AUC of93.44%,Precision of 87.04% and Recall of 80.01% on the test set.(3)In order for the features extracted from the left side contraction path in the AE U-Net network to be more accurate,the network depth can be adjusted to obtain better segmentation results.However,due to the network degradation problem,the representation ability of the deep network is not necessarily better than shallow network,and the residual module solves this problem well.AE U-Net was improved,and an AE Res U-Net is proposed.The model uses the residual network with Layers of 18 feature extraction module to replace the left contraction path in the AE U-Net,while the right expansion path keeps the original structure unchanged.The model has lower computational complexity and higher Segmentation accuracy.Experimental results on a custom breast ultrasound database show that the proposed AE Res U-Net has an M-IOU of 81.24%,Acc of93.85%,AUC of 93.44%,Precision of 85.88% and Recall of 79.48%.