Research On Intelligent Diagnosis Of Pneumonia Using Chest X-ray Based On Deep Learning

In an environment of rapid technological development,deep learning technology has been widely applied in the field of medical intelligent auxiliary diagnosis of various kinds of diseases,especially the diagnosis based on medical image.Particularly in the current environment of frequent occurrence of influenza,the pressure of health care resources has raised dramatically in the detection of diversified types of pneumonia.It is of great significance to use deep learning technology to develop deep learning model and further deploy computer aided diagnosis system to alleviate the pressure of medical care personnel and promote the effectiveness of diagnosis.Among assorted medical imaging methods,chest X-ray is more easily accepted by patients due to its portability,speed and low burden on patients’ bodies.In the process of developing deep learning models,most studies are also conducted based on chest X-ray due to the sufficient public data set.Different types of pneumonia caused by different infectious agents correspond to different treatment.Therefore,the goal of this study is to develop a deep learning model to accurately identify samples of normal,COVID-19,other viral and bacterial pneumonia from chest X-rays simultaneously.For the above pneumonia classification tasks,there are two critical points in the development of deep learning model: how to promote the pneumonia fine classification capabilities of the model for the classification tasks as far as possible,and how to avoid the problem of reduced intra-domain generalization of the trained model.The first critical point of pneumonia fine classification capabilities is the ability of the model to classify diversified types of pneumonia.The solution is generally to adjoin diversified attention modules to augment the elementary network architecture,making the model pour more attention into the areas with differences between different pneumonia image samples.The second critical point of the intra-domain generalization is the train-test generalization in the absence of external testing.The usual solution is to intervene in the training optimization process by designing the training strategy.Aiming at these two critical points,this study explored from the perspectives of network architecture optimization and model training optimization respectively,and finally proposed the multi-branch fusion auxiliary learning strategy,which was deployed and initially developed CXR pneumonia intelligent diagnosis system.The main research is as follows:(1)In terms of network architecture optimization,a dual attention module combining deformable convolution and efficient channel attention is proposed to further explore the advantages and disadvantages of single-task and multi-task network architecture modes in intelligent diagnosis of pneumonia.The former uses a network with a single output structure to perform multiclassification task directly,while the latter uses a network with multiple output structures to perform multiple correlated tasks simultaneously.Then proposed dual attention-based single-task branch network and dual attention-based task-driven multi-branch network.The results verify that the classification accuracy of the dual attention-based task-driven multi-branch network on the whole data set can reach 94.90%,which is 0.2～4.7 percentage points higher than that of single-task network.(2)In terms of model training optimization,this paper proposed the adaptive weight decay hyperparameters,the single weight penalty factor is extended to the parameter-byparameter weight penalty factor to carry out weight decay of different degrees for each parameter of the model,and adaptively adjusts the weight penalty factor based on the loss value of a small batch of online validation data independent of the training data by using gradient-based hyperparameter optimization algorithm,which called implicit differentiation.The results demonstrate that the classification performance of an elementary network can be raised from 90.25% to 95.03% on the whole data set under the intervention of the above optimization strategy,indicating that the influence of the intra-domain generalization on the pneumonia intelligent diagnosis task is more significant than the pneumonia fine classification capabilities of the model.(3)Combining the research results from the two perspectives,the following contradiction is proposed: diversified network architecture perfection used to promote the pneumonia fine classification capabilities of the model will increase the complexity simultaneously,thus amplifying the overfitting risk,so that the modified network architecture cannot give full play to its strengths,and the promoted classification performance will be partially offset by the reduced intra-domain generalization.For the sake of promoting the feature extraction capabilities of the model and keeping the complexity unchanged,the auxiliary learning strategy combined with the gradient-based hyperparameter optimization strategy was used to optimize the network architecture for the branch structure design of the auxiliary task,and then this part was regarded as the hyperparameters of the model,which was only used to assist the model training and did not participate in the model application stage after the training.By impacting model parameters through hyperparameter optimization in the training stage,the pneumonia fine classification capabilities of the model is promoted without changing the final model complexity.Moreover,the updating of hyperparameters based on the online validation data set improves the intra-domain generalization of the model.The classification accuracy of the proposed multi-branch fusion auxiliary learning strategy on the whole data set can reach 95.82%,which is 0.4～0.8 percentage points higher than that of other advanced algorithms,and the indexes in each category are better.In conclusion,the proposed multi-branch fusion auxiliary learning strategy can provide better classification performance for intelligent diagnosis of pneumonia,and can achieve the goal of rapid screening of pneumonia,providing a novel perspective for model design.