Research On CT Image Detection Of Pneumonia Based On Deep Learning

With the weakening of the pathogenicity of novel coronavirus and the opening of domestic policies,COVID-19 has gradually faded out of people’s vision.However,from the perspective of foreign countries and other regions,the COVID-19 epidemic has not disappeared.After the immune protection period,there will still be a small scale of epidemic transmission.The elderly with low immunity and underlying diseases are the first to bear the brunt,coupled with the severe aging of modern society,so the CT testing process has brought enormous pressure to local medical organizations.Manual analysis of CT images requires a high level of professional ability of doctors,and image feature analysis is time-consuming and laborious.Building a computer-aided diagnosis system has become one of the feasible solutions.In this thesis,a pneumonia CT image detection system based on deep learning is established for the field of medical image target detection.This system conducts research from both training data and model structure.It implements CT image generation technology based on generative adversarial networks and improved ground glass shadow lesion detection technology based on YOLOv5.The main work of this article is as follows:(1)In the case of insufficient training data,the unsupervised learning of convolutional generation confrontation networks is used to achieve CT image generation technology.This model uses the feature fusion pyramid and CBAM attention mechanism to jointly establish a generation network to extract and reconstruct CT images.The feature fusion pyramid only retains the maximum scale fusion,and adds an improved residual structure during the down sampling process,which relatively reduces the network complexity while improving the feature extraction ability.Finally,the convolution layer of the discriminant network is added to improve its supervision and judgment ability.On the COVID-19 dataset created by ourselves,the average detection accuracy has improved by 7.2% compared to the original YOLOv5 model,reaching 88.7%.(2)In the way of improving the network structure,this thesis proposes a ground glass shadow lesion detection algorithm based on improved YOLOv5.This method first embeds a 3D SimAM attention mechanism into backbone to improve the global feature representation ability of the network.Secondly,a C3 covid module is constructed by combining the feature processing method of the GhostBottleneck module with the residual projection PEPX to replace the original C3 module and reduce the floating-point calculation amount of the model.The above two improvements jointly build a new backbone network,CS-DarkNet53.In the neck section of the algorithm,increase the depth of FPN and PANet,establish a larger scale prediction output,and enhance the detection ability of smaller targets.Under the condition of data expansion,the average detection accuracy reaches 98.1%.(3)Design and develop the focus detection system of COVID-19 CT image.On the one hand,design a hardware system and deploy the model to the Jetson TX2 development board to achieve rapid detection.On the other hand,human-computer interaction and visualization of detection results are achieved through software systems.