Research On Chest X-ray Diseases Classification Algorithm Based On Convolutional Neural Network

Currently,the diagnosis of chest diseases usually requires experienced doctors to interpret and label Chest X-ray(CXR)images manually.However,this method of manually processing CXR images not only consumes much time and human resources,but also easily leads to misdiagnosis and missed diagnosis.Therefore,a primary current focus is exploring robust and efficient machine learning methods to detect and classify chest diseases and assist professional doctors in making diagnoses.The deep learning technique is adopted in this paper to conduct automatic and accurate classification and recognition of 14 chest diseases in CXR images.The main research work is as follows:1.Aiming at the low accuracy of the algorithms in CXR diseases classification,a Chest X-ray diseases classification algorithm based on Densely Connected Network and Dynamic Convolutions is proposed.Firstly,the Dynamic Convolution Blocks are added to the Densely Connected Network.The feature extraction capability of the network for multi-scale information is upgraded.Secondly,the Re LU activation function is improved using Meta-ACON.The non-linear expression ability of the network is enhanced.Thirdly,the weighted focal loss function is proposed.So that the proportion of losses from diseases that are difficult to classify targeted increases to improve their classification accuracy.Finally,the data loading method is optimized and a test-time data augmentation is used to increase the accuracy and robustness of the classification.The experimental results on the Chest X-ray14 dataset show that the value of the average Area Under the Curve(AUC)of the proposed classification algorithm based on Densely Connected Network and Dynamic Convolutions achieves0.8361.Meanwhile,the highest AUC value of a single disease label is up to 0.9534.The results above are preferable over a variety of advanced algorithms at home and abroad.2.To further improve the accuracy of CXR diseases classification,a scheme for classifying CXR diseases based on Dual-Mode Feature Fusion is proposed.Specifically,a novel feature extraction method is proposed based on the Scharr operator for gradient feature extraction and weighted fusion of each input image.Besides,a novel deep-learning-based feature fusion framework is proposed,which can independently learn the preprocessed features and fuse the extracted deep dual-mode features for classification accuracy improvement.In the end,the network is trained in stages through different loss functions,which can effectively alleviate the over-fitting phenomenon while accelerating the speed of network convergence.Moreover,based on Dual-Mode Feature Fusion,a Level-by-Level Feature Fusion scheme is proposed to make full use of the fused feature information.The experimental results show that the proposed two schemes achieve better accuracy of chest diseases classification,which is suitable for the classification of chest diseases in X-ray films.In particular,the highest AUC value of 14 chest diseases reaches 0.8445,and the highest AUC value of a single disease label is up to 0.9207.