X-Ray Diagnosis Of Common Chest Lesions Based On Deep Learning

X-ray has the advantages of economy and rapidity.It is an important means for early detection of pulmonary tuberculosis,pneumonia,pneumothorax and other respiratory diseases.It is known as the "flow entrance" of modern medical images.Therefore,it will be a milestone for "AI+ medical health" to benefit the public to carry out the project of screening and diagnosis of common chest diseases based on artificial intelligence by starting from X-ray,an imaging equipment covering a large number of grassroots medical and health institutions.Among them,how to improve the performance of computer-aided diagnosis based on deep learning method is one of the important problems in the field of medical image and deep learning.Focusing on key issues such as network feature learning redundancy,inadequate data imbalance,and multi-modality of case information in the fine-grained assisted diagnosis of chest X-ray images,this article proposes a method to assist common chest lesions Diagnosis,main research work and contributions include: 1.Chest X-ray assisted diagnosis based on chestx-net network Aiming at the problem of network learning redundancy in the diagnosis of X-ray chest lesions,this paper proposes a ChestX-Net network.The network,on the one hand,learn to 'feature recalibrate' in each feature channel.On the other hand,based on the Global Max-Average pool to obtain the complementary information of the feature locality,and reduce the information loss in the process of dimension reduction.The experimental results showed that the optimal AUC value of 0.813 was obtained by chestx-net network in the diagnosis of 14 lesions,which was higher than that of 0.790 by Resnet50 network and 0.799 by DenseNet121 network.Moreover,in the detection of lesion area based on weakly supervised learning,the coincidence degree with expert labeling is higher,which provides corresponding visual support and interpretability for the excellent fine-grained diagnostic performance of the network.2.Multi-label Chest X-ray Diagnosis for Imbalanced Insufficient Data Aiming at the problem of fitting difficulty and feature learning bias in multi-label diagnosis with insufficient and unbalanced disease samples and unbalanced classification of the network.This paper based on the characteristics of the gradual change of feature hierarchy in the network learning process,a discriminative collaboration based on transfer learning is propose to improve the performance of network transfer learning through feature differential learning.Then based on the multi-label focus loss function,the weight is adjusted by the focus factor to reduce the weight of simple and easily separable negative samples in the loss value,which makes the network lean to the learning of difficult and error-prone samples.Compared with the relevant work,the method in this paper obtained the 0.830 optimal AUC value in 14 kinds of common chest lesions,and verified the excellence of this method through the visual evaluation of pathological features.3.Assisted diagnosis of common chest lesions based on multiple information fusion learning In the current trend of diversification of medical information,in order to solve the problem of collaborative diagnosis of network in multi-dimensional information,we took chest lesions as the research focus,further introduced case semantic information on the basis of X-ray images,and proposed a multi-information fusion network based on deep learning.Firstly,the semantic information and X-ray image are abstracted into higher-level feature vectors through two parallel sub-networks,and then the feature fusion is carried out and the fully connected classification layer is input to achieve the diagnosis of lesions.In addition,we further studied the correlation between image data and semantic information in chestx-ray14 data set,so as to provide corresponding ideas for accurate diagnosis of network.Experiments show that the multi-information fusion network constructed in this paper is capable of joint cross-modal relationship modeling and multi-feature complementary learning.With the opening of the wave of case information diversification,it will have great development potential in clinical practice.