The Research Of X-ray Image Analysis Based On Deep Convolutional Neural Network

With the rapid development of digital medical imaging,X-rays have become an indispensable technology for modern hospitals.Wrist bone X-rays are used to diagnose children's growth problems,while anterior pelvic X-rays are used to detect femoral neck fractures in middle-aged and elderly people.The lesion information of these two X-rays requires experienced radiologists to spend more than a dozen minute screening can be accurately detected.Traditional computer-aided diagnostic algorithms use artificially designed features to assist doctors in quickly extracting lesion information from X-rays.In recent years,deep learning has carried out extensive and in-depth research work on medical images,eliminating the complicated manual extraction of medical image features.However,due to the complexity of the lesion information and background information in X-ray images,general deep learning methods cannot effectively extract features in X-ray images.Our study focuses on researching deep learning methods to extract effective information in different medical X-ray tasks.Therefore,our study proposes an improvement plan for the bone age assessment task and the femoral neck detection task,mainly including the following:In the bone age assessment task,aiming at the problems of wrist bone X-ray lesions with small differences,difficult to distinguish and multi-source heterogeneous data performance in bone age assessment task,this thesis proposes a new fusion learning network RT-FuseNet.RT-FuseNet includes two branches,one is an image sub-network for learning image features,and the other is a text sub-network for learning text features.In the image sub-network,the spatial pyramid pooling layer is used to avoid the loss of the information in the X-ray image caused by the pooling operation in the general network.Secondly,in image sub-network,channel attention and spatial attention are added to enhance the distinguishability of key areas in X-ray images.RT-FuseNet evaluates the bone age of the wrist X-ray by concatenating the features of the two sub-networks.The validation on the public dataset that comes from the 2017 Pediatric Bone Age Challenge organized by the Radiological Society of North America(RSNA)and a private dataset show that the proposed network based on multi-source heterogeneous data in this thesis can learn more effectively features,resulting in a better performance.In the anterior pelvic X-ray femoral neck fracture task,aiming at the complexity of femoral neck fracture information and background context information of anterior pelvic X-ray images,it is prone to the problem of extraction of wrong features and model overfitting.Our study proposes a new cascaded deep learning model.Firstly,this study trained a fully supervised detection neural network to exploit the femoral neck area in the anterior pelvic X-ray,avoiding the feature of extraction errors due to the complexity of fracture information and background information in the anterior pelvic X-ray;Secondly,our study compares the performance of VGG,ResNet,and DenseNet architectures in classification tasks,selects the DenseNet architecture as the network skeleton of the semantic segmentation algorithm,and constructs a U-net-like semantic segmentation model to visualize fracture fractures.The feasibility of the proposed method was verified in the private data set of the anterior pelvic X-ray.Finally,aiming at the problems of uneven distribution of medical resources and people's increasing medical needs,the study designs an X-ray image analysis system based on deep learning.It mainly includes data storage module,data preprocessing module,bone age assessment module and femoral neck analysis module.After testing,the system can normally meet the needs of users.