Research On Recognition Of Modal Features Of New Coronary Pneumonia Images Based On Deep Learning

COVID-19 has had a huge impact on the existing medical system in the past year.In many areas abroad,the medical system has exposed the weaknesses in the face of major epidemics,the shortage of medical equipment,and the shortage of medical staff.At present,the diagnosis of COVID-19 is still heavily dependent on antibody detection reagents and the diagnosis of lung CT images.With the development of deep learning,the recognition of modal features of new coronary pneumonia images based on deep learning can assist the diagnosis of new coronary pneumonia based on lung CT images..However,CT images of the lungs have quite complex features and relatively few data sets.How to use the limited data set to train a model that can identify new coronary pneumonia and train a new coronary pneumonia target detection model that can detect the focus of the disease is the focus of this experiment.In response to this problem,I based on the deep learning image recognition multi-training model fusion,based on YOLO for the secondary training of new coronary pneumonia target detection,and conducted some research in these two fields.My main work is:(1)A series of image augmentation methods have been adopted for the large demand for new coronary pneumonia COVID-19 image recognition samples,and the recognition accuracy is not ideal,such as the use of data enhancement for colors,PCA dithering,scale transformation,and the use of random image differences to augment Image and pseudo-color processing of the image to generate a new data set,which significantly improves the recognition accuracy of target detection.At the same time,because of the large number of images,the training has sufficient samples,avoiding the appearance of over-fitting,and making the model have better generalization ability.(2)In response to the need for rapid isolation of patients with new coronary pneumonia COVID-19,a model for rapid recognition of positive images of COVID-19 was trained this time.This model uses residual connections to forcefully break the symmetry of the network,reduce the degradation of the neural network,and increase The detailed network extraction layer integrates some existing models through model migration.The experimental results show that the final training model has a high recognition accuracy for the classification and recognition of new coronary pneumonia positive images.This model is a two-class model with high accuracy and fast model reasoning ability,which can screen patients for the first time,which greatly reduces the burden on medical staff.(3)In response to the shortage of medical staff,a model for detecting lesions in COVID-19 positive images was also trained this time for the diagnosis of COVID-19 patients with new coronary pneumonia.The model uses opencv to color-process the image,and superimpose the processed image,and then train the pseudo-color image.The experiment also performed a special optimization of the YOLO algorithm for pneumonia-positive images.Through migration The mature YOLO V3 model is used to conduct secondary training on the pseudo-color new coronary pneumonia images.Experimental results show that the COVID-19 positive image image lesion detection model can better detect the lesion area.The detection of lesions by target detection can perform a second screening of patients,which has certain guiding significance for medical staff in the diagnosis of new coronary pneumonia.Of course,antibody testing reagents will eventually be used to confirm whether it is a COVID-19 patient.