Recognition And Research Of Pneumonia Based On Convolution Neural Network

Pneumonia is one of the most common infectious diseases in the clinic.Due to its short onset period and complex causes,the number of children under the age of 5 is more than 800,000 deaths per year due to pneumonia-related diseases worldwide.Early identification plays an important role in the diagnosis and post-treatment of the disease.X-ray examination has been widely used in imaging examinations in various regions because of its advantages in radiation,price and practicability.Most of the chest lesion information can be obtained by observing X-ray films.However,there are some shortcomings in X-ray images.For example,the image resolution is low,and the overlapping of the organs in the image leads to the fact that the true lesion position is easily covered,which is difficult to distinguish.The doctor is easy to misdiagnose and misjudge,which increases the difficulty of diagnosis.Computer-aided diagnosis of lung diseases has become one of the important topics in the medical field in recent years.The main research content of this paper is the identification and research of pneumonia based on convolution neural network.Around this content,the following work was mainly carried out:(1)Lung field segmentation based on convolution neural network(CNN)Firstly,the X-ray chest image is preprocessed.After filtering operation and Gauss-Laplace pyramid enhancement algorithm,the average contrast of the image is greatly improved compared with the original image.Secondly,aiming at some difficulties in X-ray chest image segmentation,combined with the related techniques of deep learning,a lung field segmentation method based on CNN is proposed.This method is a feature extraction using the super-pixel image block in the lung region.This convolution neural network model based on pixel segmentation densely divides the chest X-ray.In this paper,by improving a fully connected layer with multiple parameters,it is converted into a convolution kernel with a convolution window of 1*1,which enables the network to efficiently perform block-by-block training,thus improving the convergence of network training.By comparing the traditional Ostu segmentation algorithm,the general CNN algorithm and the improved CNN segmentation algorithm,the experimental results show that the lung field contours segmented by the first two methods have a slight Dice coefficient of only 0.792 due to the slight influence of the rib region.And 0.893,and the method of this paper accurately segmented the outline of the lung field,and the Dice coefficient of the segmentation performance index reached 0.948.(2)Pneumonia recognition based on convolution neural network(CNN)The pneumonia identification part first briefly introduces the technical route of pneumonia recognition,the general process of network training,and then uses the classic VGG-16 network to train the unamplified data set to obtain preliminary experimental results.This paper uses sensitivity and specificity.Evaluation indicators such as degree and accuracy are used to evaluate the performance of the model.Secondly,combined with the characteristics of deep learning,the following improvements are made: including optimization of network structure,changing the number of network layers,reducing the number of convolution layers,and fine-tuning the hyperparameters of prototxt files in network training,thereby improving the convergence speed and generalization ability of the network model.Finally,the 3-fold cross-validation of the amplified dataset was carried out,and the experimental results were analyzed and compared with other methods.The results show that the proposed VGG-9-LRN model is better than the other pneumonia recognition algorithms listed in the paper,and the recognition accuracy can reach 96.3%.