Research On Pneumonia Detection Algorithm Of Chest X-rays Based On Deep Learning

Pneumonia is one of the leading causes of death for children and the elderly around the world.More than 15% of deaths worldwide are caused by pneumonia,including children under 5 years of age.Pneumonia is an infection of the lungs caused by bacteria,viruses,or fungi,causing inflammation of the lungs and causing alveoli to fill with fluid such as pus.If not diagnosed in time,it may be life-threatening.Chest X-ray is an important diagnostic method for pneumonia worldwide.However,an expert with knowledge and experience is required to carefully read the chest X-ray image.The chest X-ray image is an important basis for detecting chest diseases such as pneumonia.However,since several other diseases of the chest such as lung cancer and excessive fluid in the lungs will also display visual signals similar to pneumonia on the image,the chest X is manually read.The process of detecting pneumonia using radiographic images can be time-consuming and less accurate.Therefore,in order to better assist the radiologist to read the chest X-ray image,it is very meaningful to develop a model for automatically detecting pneumonia.At present,the pneumonia detection model faces two problems: First,the detection of pneumonia in chest radiographs is different from the traditional target detection.In the traditional target detection,the characteristics of the target are very obvious,but the characteristics of the visual signal of pneumonia in chest radiographs are not obvious;Second,for the detection of pneumonia,not only need to observe chest radiographs,but also ask the patient’s clinical history.These problems have brought certain difficulties to the detection of pneumonia.Recently,with the development of artificial intelligence,the application of deep learning to medical images has shown great advantages.With the continuous development of deep learning in target detection and the realization of high performance,the problem of pneumonia detection can be better solved.Deep convolutional neural networks have achieved great success in target detection.Based on the advantages of deep learning,this paper applies deep convolutional neural networks to pneumonia detection.The main research work is as follows:(1)The main problems facing pneumonia detection are studied.First,the characteristics of the target area of the chest X-ray image are not obvious.The visual signals of other diseases of the lung are similar to the visual signals of pneumonia.The traditional image target detection method is difficult to detect the pneumonia area.Second,the target area of pneumonia is small,which will cause difficulties in detection.Third,the vast majority of pneumonia data sets are a data set with unbalanced positive and negative samples.Although the total data volume is large,the data with value is relatively limited.(2)For the problems mentioned in question(1),this paper applies the deep convolutional neural network RetinaNet to the research work of pneumonia detection.The fusion of pneumonia detection models based on two different feature extraction networks based on ResNet-50 and ResNet-101 was studied.Since the position of most pneumonia is relatively small,in order to improve the detection accuracy of the target position,this paper has made certain modifications to the network structure on the basis of the RetinaNet network,using a deeper FPN network,respectively established based on Pneumonia detection models of Res Net-50 and ResNet-101 two different feature extraction networks,and then the RetinaNet models obtained under the two different feature extraction networks are fused.The experimental results show that the mAP and Recall indicators of the model after fusion have reached the effects of 0.2196 and 0.796,respectively,which are higher than the detection effect of the single model,which is 0.0259 and 0.017 higher than the model based on ResNet-50 and higher than the model based on ResNet-101 0.037 and 0.031 higher.(3)The fusion-based Mask R-CNN pneumonia detection model is studied based on the fusion of pneumonia detection models under ResNet-50 and ResNet-101 two different feature extraction networks.Mask R-CNN is a deep neural network developed to solve the problem of instance segmentation.Mask R-CNN is a deep neural network developed to solve the problem of instance segmentation.In view of its advantages in target detection and segmentation,based on the Mask R-CNN network,this paper still uses two different feature extraction networks ResNet-50 and ResNet-101 and a deeper FPN network.Mask R-CNN is a deep neural network developed to solve the problem of instance segmentation.In view of its advantages in target detection and segmentation,based on the Mask R-CNN network,this paper still uses two different feature extraction networks ResNet-50 and ResNet-101 and a deeper FPN network,and then two different features The Mask R-CNN model obtained under the extraction network is fused,and after the fusion,the mAP and Recall achieve the effects of 0.2102 and 0.760,which are improved compared to the Mask R-CNN itself.It is 0.0162 and 0.004 higher than the model based on ResNet-50,and 0.02 and 0.015 higher than the model based on ResNet-101.Through the research content(2)and(3),it can be seen that the fusion of the pneumonia detection model based on the two different feature extraction networks proposed in this paper has significantly improved the effect of pneumonia detection,which is Because the bounding box adjustment strategy with confidence as the weight helps to improve the effectiveness of pneumonia detection.(4)A pneumonia detection model based on the fusion of Retina Net and Mask R-CNN is proposed.The model fuses two different network pneumonia detection models.According to the confidence as a weight,the two different pneumonia detection models are fused in a weighted average.Since the single model RetinaNet performs best,the RetinaNet pneumonia detection model is the main as the auxiliary pneumonia detection model,the Mask R-CNN pneumonia detection model is used to modify the prediction of the RetinaNet pneumonia detection model.The experimental results show that the pneumonia detection effect after the fusion of two different network models has achieved excellent results of 0.2283 and 0.813 on the test set,which are higher than the single model and single model fusion methods.(5)In order to further illustrate the reliability of the algorithm proposed in this paper,it is compared with classical Faster R-CNN network and YOLOv3 and other algorithms.The performance of mAP and Recall on the test set are 0.1573,0.618 and 0.1647,0.703 respectively.We have improved on the structure of Faster R-CNN and YOLOv3 network,but the detection is due to the uneven distribution of data and the small target area.The effect is much lower than the above methods.It can be seen that the algorithm of pneumonia detection model based on the fusion of RetinaNet and Mask R-CNN proposed in this paper is more reliable and the detection effect is better.