Study Of Pediatric Pneumonia Assisted Diagnosis Algorithm Based On Deep Learning

The symptoms and signs of pneumonia in children are not very obvious at an early stage,and the disease is easily confused with other diseases.Therefore,the rapid detection and diagnosis of pneumonia have important significance in improving the survival rate and quality of life of children.As a common detection method for pediatric pneumonia,chest X-ray examination can help doctors observe the structure and organs of the chest.However,due to the unbalanced distribution of medical resources,the experience and knowledge reserve of pediatricians in hospitals at all levels are quite different.In addition,it is difficult for young children under the age of five to consciously cooperate with the doctor’s instructions when performing chest radiographs.Compared with adult chest X-ray imaging,its imaging quality will be relatively poor.These factors have a very negative impact on the reliability and stability of current manual diagnostic results.In view of this,the development of a computer-aided diagnosis system for pneumonia based on chest X-ray is particularly important for sharing the work pressure of pediatricians and improving the efficiency and quality of clinical medical image analysis.With the continuous progress of deep learning algorithms in the field of computer vision,researchers have begun to extend this artificial intelligence algorithm to the field of medical image analysis.However,most of the deep learning algorithms only focus on the improvement of accuracy,and ignore the limitations of computing power and memory space of medical detection equipment.This brings great challenges to the productive application of deep learning related technologies in the field of medical diagnosis.Based on the above problems,this thesis studies and proposes a small-scale deep learning algorithm for pediatric pneumonia assisted diagnosis.This algorithm was used to classify children X-ray chest radiographs into pneumonia images or normal images.Specific contributions and innovations include:(1)From the perspective of expanding the receptive field in the depth model,this thesis proposes a dilated convolutional neural network model(DCNET)for aided diagnosis of pediatric pneumonia.In the model design process,this thesis uses dilated convolutions instead of ordinary convolutional layers,and extracts features from the pixel-level raw data layer by layer by stacking many convolutional layers to generate a classification result representation.In this way,a wider range of input views can be obtained without increasing the computational complexity of the model,which enables the model to better aggregate image context information.The experimental results show that the proposed DCNET network has a classification accuracy of 89.10% with a model size of 2.08 MB.(2)From the perspective of enhancing the discrimination ability of feature maps,this thesis proposes a method that combines feature fusion with attention mechanism to improve the structure of DCNET network to improve its classification performance.The improved network is named RES-SE-DCNET.The proposed method enhances the model’s learning of low-level detail features by multi-level feature fusion,then introduces an attention mechanism to make the model better focus on the feature information related to the target task,and finally make the features learned by the model have stronger discrimination.The experimental results show that the RES-SE-DCNET model designed here has a size of 2.09 MB,and the classification accuracy on the test sample is further improved to 89.74% with only a little parameter added.(3)From the perspective of network parameter initialization,this thesis proposes to use the transfer learning method to further improve the classification performance of the RES-SE-DCNET network.Different from the traditional method of pre-training the model using the ImageNet dataset,this thesis uses transfer learning techniques to initialize the RES-SE-DCNET network using weight parameters learned on other largescale chest X-ray image datasets.The experimental results show that this transfer learning-based optimization method can effectively solve the impact of relatively insufficient data on model performance,and the classification accuracy on the test sample is further improved to 90.71%.The classification method of each chapter in this thesis is improved based on the method of the previous chapter,and the algorithms are organically linked.Each improvement to the classification method has resulted in better classification performance.