Deep Neural Network-Based Feature Learning Methods For Medical Image Analysis

Medical image analysis is a multidisciplinary field with many techniques,including medical imaging technology,mathematical modeling methods,artificial intelligence technology,et al.A typical method of medical image analysis mainly consists of several key processes including digital image processing,feature analysis of images,and decision-making.In particular,the quality of the feature analysis affects the performance limit of the method greatly.Traditional methods in medical image analysis mostly rely on manual feature engineering technology,requiring experts to spend a lot of time and effort designing features manually with corresponding prior knowledge for specific medical tasks.However,the features designed for specific scenes usually lack generality,and their adequacy and fineness are also limited.As a result,the current mainstream study direction of feature analysis methods in medical image analysis has started to shift from feature design to feature learning.Deep learning methods,especially deep neural networks,can be used to perform direct feature learning from medical images automatically and implicitly,achieving excellent performance in a variety of tasks in medical image analysis.The annotations of medical images are scarce and the characteristics of medical images are complex,requiring more powerful feature learning methods based on deep neural networks.To meet this challenge,the author analyzed the current problems in medical image analysis and tried to solve them by means of the deep neural network-based feature learning methods.In this thesis,the author proposed several novel network frameworks for diverse feature extraction,few-shot feature expression,and multi-level feature fusion in medical images and applied the methods in the applications of medical image analysis.The main works of this thesis are summarized as follows:1.The methods of diversified feature extraction on medical images.Medical images contain information about human organs and tissues,exhibiting topological complexity in structural shape and distribution.Existing medical image processing methods based on deep neural networks tend to focus only on a single description of the local or global information of images,resulting in weak recognition performance in complex scenes.In order to enhance the ability of feature learning methods in accurate and rich feature extraction,the author designs a composite network framework,which combined a multi-scale convolutional neural network and a graph convolutional neural network as the backbone network,for diversified feature extraction.On the basis of that,the author proposed and evaluated a disease identification model for the detection of abnormalities in the Musculoskeletal Radiographs(MURA)dataset.The accuracy of the model was 91.22%,and the F1 score and the Kappa score were 0.909 and 0.836,respectively.The results show that the proposed diversified feature extraction method can obtain more accurate,fine,and comprehensive features from medical images of high complexity and improve the comprehensive performance of the disease recognition model.2.Feature expression methods in medical images from few samples.In the actual medical scene,large-scale labeled medical image data are scarce,showing the characteristics of "few samples".However,the current deep neural networks are mostly supervised learning methods,and the training process relies on large-scale and high-quality labeled data.To improve the feature representation capability of feature learning methods in medical image analysis under few samples conditions,thus enhancing the comprehensive performance of analysis models.The author proposed a semi-supervised learning framework consisting of a two-branch network structure with an attention mechanism.The framework,used to complete feature learning from medical images,can learn the underlying feature expressions from unlabeled data and the high-level feature expressions from labeled data,respectively.The proposed semi-supervised learning framework was evaluated on two public brain MRI datasets,the Kaggle Alzheimer Classification Dataset(KACD)and the Recognition of Alzheimer Dataset(ROAD).The accuracy of the framework on these two datasets were 0.9961 and 0.9871,respectively,which is not only better than that of the supervised learning method which only learns from a small amount of labeled data,but also 2.50%and 1.88%higher than the other two advanced semi-supervised learning framework(ResNeXt WSL,SimCLR).The results show that the proposed semi-supervised learning framework has a potential in solving the "low precision" problem of disease identification model based on deep neural network under "few samples" conditions.3.Medical image analysis based on multi-level feature fusion.Medical images have rich information including underlying pixels,mid-level objects,and high-level semantics et al.,of which the high-level semantic information is often closely related to image understanding.Existing feature learning methods of medical images usually only focus on information mining on the pixel or object levels,with less focus on medical image analysis based on multi-level feature fusion.The author designs a multi-layered and modular convolutional neural network framework to achieve bottom-up feature learning by introducing a control gate module.The framework accomplishes feature fusion between different levels in a deep network,so as to strengthen the multi-level feature representation capability.The proposed framework was evaluated on two public chest X-ray datasets(CCD and RSNA),two public computed tomography image datasets(LUNA 16 and ICNP)and clinical image data from Beijing You’an Hospital(95 cases).The results show that the proposed framework achieves a performance close to that of experts in the detection of COVID-19 disease.The accuracy,sensitivity and specificity of the framework are 98.33%,95.16%and 99.33%respectively.Besides,the proposed framework also demonstrates the ability of mapping the relationship between the visual features and the clinical indicators which provides support for the correct analysis of medical images and causal inference in clinical medicine applications.4.Applications of computer-aided diagnosis utilizing the proposed methods.Computer-aided diagnosis,a typical application scene of medical image analysis methods,requires high inference efficiency and accuracy.Based on the previously mentioned feature extraction,expression,and understanding methods,the author proposed further improvements and optimizations in the data processing method and the model structure design to meet the requirements of computer-aided diagnosis.The author proposed a data resampling method and designs a composite network framework.The methods achieved an F1 score of 88.21%in the diagnosis task of COVID-19 detection on the COV19-CT-DB dataset,which outperformed the baseline method by 18 percentage.A multi-level modular network model is constructed using a two-branch network framework.The framework consists of a deep convolutional network and a deep self-attentive transform network to extract spatio-temporal features simultaneously from abdominal computed tomography images.The proposed framework achieved 97.56%grading accuracy and 92.54%Dice score on the 2019 Kidney and Kidney Tumor Segmentation Challenge(KiTs19).The framework’s classification accuracy is 6.25%and 9.38%higher than the other two advanced models(ResNeXt,ViT),and its segmentation performance is 1.36%and 1.80%higher than the other two advanced models(Hybrid V-NET,Unet 3+),respectively.The proposed methods and innovative network structures can provide theoretical and technical support for applications of medical image analysis technology in intelligent healthcare.This study has positive social value and potential commercial value,which can help accelerate the progress of deep learning methods in medical image analysis and clinical medical practice.