Research On Intelligent Method For Local Feature Representation-based Mammogram Diagnosis

According to the studies conducted by International Agency for Research on Cancer,breast cancer has surpassed lung cancer as the most commonly diagnosed cancer worldwide in 2020.The number of new cancer cases in China is approximately 4.57 million in 2020,of which more than 410,000 are breast cancer patients,accounting for 9.1%of the total number of new cancer cases in China.As it enhances breast cancer survival effectively,early detection has been proven to be the cornerstone of breast cancer control.Since it is a non-invasive detection method and is able to pinpoint tiny abnormalities before being felt physically,mammography has been widely used to detect breast cancer at an early stage.With the popularization of mammography screening,the following problems have arisen:(1)Compared with the huge number of early screen people,the number of experienced radiologists is very limited;(2)Due to the unclear characteristics of lesions and the subjectivity of diagnosis,radiologists are unable to give consistent conclusions sometimes;(3)Compared with developed areas,medical conditions in remote areas are backward and the level of doctors is generally lower.These factors restrict the efficiency and accuracy of doctors’ diagnosis of breast cancer through mammograms.Developing automated and robust computer-aided diagnosis approaches could increase the detection rate and improve the accuracy of diagnosis through mammograms.Traditional machine learning methods that rely on hand-crafted features from lesions must be designed according to specific data meticulously.Furthermore,these methods may have poor portability because hand-crafted features are not data-driven.Deep neural networks(DNNs)can automatically extract features from raw data in a data-driven manner,and have powerful feature extraction capabilities,feature representation capabilities,and feature generalization capabilities.Several DNNs-based approaches have been explored to screen breast cancer by mammograms.However,unlike natural images,the lesions in mammograms are variable in shape,and tiny in size,the local features are greatly affected by global features,and multi-view data are highly correlated and complementary.Those existing assisted diagnosis methods do not fully take into account the characteristics of mammograms and also ignore the diagnostic process of clinicians.This thesis conducts research on three key stages for automatic mammogram diagnosis,namely,whole mammogram classification method,multi-instance mammogram classification method,and mammogram weakly supervised object localization method.The innovations and main contributions of this thesis are as follows:1.Adaptive convolutional feature descriptor selection moduleBased on the characteristics of mammograms,this thesis designs a deep learning method for classifying benign and malignant mammograms.For this method,this study designs a novel adaptive convolutional feature descriptor selection module to select discriminative feature descriptors(lesion areas)adaptively.Ablation experiments and visualization analysis indicate that the proposed method makes the model easier to learn the difference between malignant and benign/normal lesions.Experimental results on two publicly available mammogram datasets indicate that the proposed method performs satisfactorily compared with state-of-the-art methods.2.Deep multi-instance classification method with region label assignment strategyConvolutional neural network-based methods have obtained promising results for automatic mammogram diagnosis.However,most of those approaches ignore the relationship between global and local characteristics of mammograms and lose sight of the relationship between different views of a patient.This thesis designs a novel region label assignment strategy,which takes advantage of all regions in each mammogram from a patient by assigning different labels to different regions and calculating loss for each region separately.This approach enables the classifier to distinguish variable and tiny lesions in complex global conditions better.And only one case-level classification label is needed for diagnosing one patient(case).Moreover,the categories of mammogram datasets are always imbalanced.To address this problem,this thesis designs an AUC(area under the receiver operating characteristic curve)-based optimization method on mini-batch strategy.Experimental results on a constructed dataset and two publicly available mammogram datasets demonstrate that the proposed method performs satisfactorily compared with state-of-the-art mammogram classifiers.3.Long-short feature fusion method for weakly supervised object localization of mammogramsBy combining the sensitivity of convolutional neural networks to local information and the advantages of Vision Transformer networks in long-range semantic relevance,this thesis proposes a novel long-short feature fusion method for weakly supervised object localization of mammograms.The method first extracts the feature map from the convolutional neural network branch,and then couples the feature map of the convolutional neural network branch into the initial block embedding of the vision transformer branch through the feature fusion module.Subsequently,this method feds the fused features into the vision transformer network for training.In order to make full use of the long-range context dependency of the transformer,this method aggregates the attention matrices of all transformer blocks to generate an attention map.At the same time,considering that the lesion area in mammograms usually occupies a relatively small area,this thesis designed a size constraint loss function to make the located lesions more consistent with the characteristics of mammograms.Experimental results show that the proposed method can achieve state-of-the-art weakly supervised object localization performance on public mammogram datasets.4.Contrastive learning-based class-agnostic dual-branch weakly supervised object localization methodIn order to further generalize the weakly supervised localization method to natural image datasets,this thesis proposes a novel contrastive Learning-based class-agnostic dualbranch weakly supervised object localization method.This method first interactively fuses local features generated by convolutional neural networks and global features generated by vision transformer layer by layer.Then this method uses a novel clusteringbased cross-image foreground-background contrastive learning loss function to force the model to decouple foreground and background of images.Experimental results on two benchmark datasets verify the performance and robustness of the proposed method.In summary,this thesis focuses on the problems and challenges in the field of mammogram intelligent diagnosis and carries out algorithm research from the aspects of whole mammogram classification,multi-instance mammogram classification,and weakly supervised object localization of mammograms,based on local feature representation.It improves the accuracy and interpretability of the intelligent diagnosis method of mammograms and also provides effective technical support for the research of intelligent analysis of mammograms towards clinical practice.