Breast Ultrasound Image Tumor Segmentation Based On Weakly Supervised Learning

As a leading cause of death for women worldwide,breast cancer has attracted extensive attention.Early detection of breast cancer can significantly improve survival.With the advantages of no ionizing radiation effects,low price,and so on,ultrasound has been recognized as a universal and effective screening diagnostic tool for clinicians and radiologists.The appearance of computer-aided diagnosis system significantly improves the objectivity and accuracy of diagnosis.Breast segmentation based on deep learning has developed rapidly,but there are still some problems.Due to the imaging characteristics of breast ultrasound,the image contrast and resolution are low,the speckle noise is high,and the boundary between tissues is fuzzy.Therefore,this thesis attempts to build a suitable network structure to achieve accurate breast tumor segmentation under noise annotation.Firstly,the dynamic noise index(DNI)and segmentation loss function are proposed to detect the noise data.Noise index is used to represent the probability that the sample is noise,and it is updated dynamically in the training process.An improved loss function is proposed to enhance the performance of network noise detection and reduce the influence of noise on training back propagation.Finally,high performance segmentation is achieved when the breast ultrasound data contains part of the noise.Secondly,for the utilization of noise data,this thesis proposes a noise tolerance network based on weakly supervised(NAT-Net).The network is divided into two parts: segmentation module and correction module.The segmentation module uses noise index to detect noise and uses convolution block with multi ratio to extract feature map;the correction module dynamically corrects noise annotation by combining feature map and model prediction results.NAT-Net is superior to the existing methods in data with different noise ratios.For real datasets with more complex noise types,the Io U of NAT-Net is nearly 6% higher than other methods.Based on the clean data set,NAT-Net also achieves competitive results.Finally,an activation induced circle network is proposed to segment and classify breast tumors simultaneously.The activation map is obtained from the classification network,which provides prior guidance for the segmentation network.The location map generated by the segmentation model transfers the edge information and morphological information of tumor to the classification network to achieve more accurate classification of benign and malignant tumors.I n this way,the segmentation and classification network can transfer the information of tumor lesions and promote each other in a guided way.Activation induced circle network reduces the manual interaction of data annotation and the dependence on medical staff.After tumor segmentation,the efficiency of disease diagnosis is improved,which provides an auxiliary strategy for medical diagnosis system based on ultrasound.