Research On Pneumonia Assisted Diagnosis Of Chest X-rays Based On Deep Learning

Pneumonia is one of the serious diseases that cause the death of children,and the novel coronavirus pneumonia is widely prevalent all over the world,so the rapid detection of pneumonia is of great significance.Pneumonia detection not only requires high detection accuracy,but also requires good real-time performance.Because pneumonia is not well characterized in medical images,radiologists spend a lot of time diagnosing the presence of pneumonia in chest X-rays.Watching chest X-rays for a long time may cause the doctor’s vision to become fatigued,which may lead to missed and misdiagnosed pneumonia in chest X-rays.With the development of artificial intelligence,the use of deep learning methods to achieve automatic detection of pneumonia and help radiologists to assist in the diagnosis of pneumonia in chest X-rays has certain practical significance.The work of this paper mainly includes the following four aspects:(1)In view of the problem of the small amount of data in the pneumonia dataset,this paper expands the pneumonia dataset through data augmentation.In order to speed up the convergence speed of the neural network,this paper uses the pre-trained model trained on the Image Net dataset to perform migration learning;in order to enable the neural network to perform normal training,this paper scales the input image.(2)In view of the problem that the number of negative samples in the pneumonia data set is much larger than the number of positive samples,this paper proposes to use Darknet53 as the basic network to achieve the purpose of balancing the positive and negative samples of the pneumonia data set by modifying the loss function.This algorithm uses the Focal loss function,and by adjusting the weighting factor and modulation factor of the Focal loss function,it effectively balances the positive and negative samples and difficult samples of the pneumonia dataset.In the Kaggle pneumonia dataset experiment,the detection accuracy AP50 of pneumonia in chest X-rays is 53.0%,and the detection speed is 32.6 frame/s.(3)In view of the problem that the deployment of mainstream deep learning object detection algorithm models requires high hardware equipment,this paper proposes to use the lightweight neural network structure Mobilenetv3 network to replace the VGG16 network in the SSD object detection algorithm.The algorithm modifies its feature extraction network by adding the optimized RFB and RFB-s network structures,and combines the idea of feature fusion on this basis,which effectively improves the object detection accuracy.Since the object detection algorithm adopts a lightweight neural network,the model can be deployed to the mobile terminal.In the Kaggle pneumonia dataset experiment,the detection accuracy AP50 of pneumonia in chest X-rays is 53.6%,and the detection speed is 26.5frame/s.(4)In view of the problems of high computational cost and poor detection effect for small objects in the Resnet101 network structure,this paper proposes an improved Resnet101 network,and by adding a dual attention mechanism CBAM module and improving the original feature pyramid structure,to improve Accuracy and computational efficiency of object detection algorithms.In the Kaggle pneumonia dataset experiment,the detection accuracy AP50 of pneumonia in chest X-rays is 63.8%,and the detection speed is10.2 frame/s.