Research Of Deep Learning For Breast Ultrasound Computer-aided Diagnosis

Breast cancer is one of the most common forms of cancer among women worldwide and has brought a great threat to women’s health.Because of the complex etiology that makes it difficult for the medical community to provide effective preventive measures,early detection and diagnosis are the keys to reducing the death rate.In recent years,more and more attention has been paid to ultrasound imaging because it has the following merits: no radiation,faster imaging,higher sensitivity and accuracy,and lower cost.However,ultrasound is an operator-dependent method.Moreover,reading ultrasound images requires well-trained and experienced radiologists and the diagnostic results of physicians with different medical attainments are very different.To reduce the inter-observer variation among different radiologists and help generate more reliable and accurate diagnostic conclusions,the computer-aided diagnosis(CAD)system has been proposed.With the rapid development of computer technology and artificial intelligence,the CAD has become one of the hotspots in modern medical image research,and its clinical value has been demonstrated in real diagnoses to assist radiologists in breast tumor classification and recognition.This dissertation takes the differentiation between benign and malignant tumors as the research content,combines the deep learning method with the characteristics of clinical data,and aims how to extract more useful features for classification.This dissertation mainly solves three problems in breast ultrasound CAD.The first problem is how to combine segmentation and classification and add segmentation prior knowledge to the classification network.The second problem is how to fuse image data and clinical omics data at two different semantic levels.The third problem is how to solve the problem of the tumor-centered image(TCI)selection in breast ultrasound classification and extract finer-grained features to get better classification results.The main contributions and innovations of this dissertation are as follows:(1)The existing breast tumor classification methods based on deep learning ignore the correlation between image segmentation and classification,and it is difficult to extract the most representative features in a small amount dataset.To solve this problem,the dissertation proposes a segmented-based attention network,which adds the segmented prior knowledge into the design of a deep convolution network for the breast ultrasound classification.The study applies a segmentation network to obtain the tumor region and edge information and constructs the segmented enhanced image.After that,we extract the features of the original and segmented enhanced image through two parallel feature extraction branches and propose a feature aggregation network based on channel attention to enhance features useful for classification and get a more effective feature representation.The experimental results demonstrate that the segmentation-to-classification framework can improve diagnostic performance and achieves higher accuracy and AUC.(2)Because the ultrasound image and BIRADS data are at different semantic levels in the process of breast tumor diagnosis,the existing breast CAD is difficult to combine their information for the breast ultrasound classification.In this dissertation,a deep learning network based on feature fusion of different semantic levels is proposed.In this scheme,the low-level semantic image and high-level semantic medical clinical BIRADS score data are paired and input into the network for the classification.Two different neural networks are proposed to map the image and BIRADS information to the same high-dimensional space.To automatically extract the features of key regions on the image,a convolution neural network method combined with a spatial attention mechanism is designed for feature extraction of the breast image.At the same time,this method also proposes an image-BIRADS feature fusion module based on the channel attention method,which integrates image features and BIRADS features to classify benign and malignant tumors.Experiments show that the proposed method greatly improves the classification performance compared with the image-based and BIRADS-based methods,respectively.(3)Aiming at the selection of TCI and fine-grained feature extraction in the process of breast ultrasound classification,this dissertation proposes a multi-view learning method for breast tumor classification.This method proposes a novel strategy,which generates multiple multi-resolution TCIs with different expansion rates in a single ultrasound image,regards each TCI as a view of the tumor,and transforms the traditional learning task based on a single image into a multi-view learning task.At the same time,the dissertation also proposes a homologous bilinear network to extract the second-order information of the feature map obtained from the backbone network to obtain a more fine-grained feature representation of the tumor.Finally,the study proposes a combined-style multi-view fusion method suitable for deep networks,which integrates the features and decision information of each view to identify the benign and malignant breast tumors.The experimental results demonstrate that compared with the existing methods,the proposed method greatly improves the classification performance.