Pulmonary Lesions Recognition And Diagnosis Based On Deep Learning

With the development of artificial intelligence and the popularization of medical images,medical images have been gradually applied in disease diagnosis.In the traditional disease diagnosis,the diagnosis is completed based on the direct experience of doctors through the characterization of patients and the analysis of relevant medical images,which is highly subjective and uncertain.According to statistics,the number of medical images is increasing at an exponential rate every day,which makes doctors face great pressure to read films.Besides,China has the problem of uneven distribution of medical resources.To solve the problems that existed in the intelligent diagnosis,this paper takes the lungs X-ray image as the research object and targets around disease intelligent technology of deep research on the important and difficult issues,such as the image super resolution reconstruction,the classification and segmentation of the image.Finally,14 kinds of pulmonary diseases can be diagnosed intelligently.The main research work includes the following aspects.(1)To solve the problems of unclear edges and fuzzy details in the existing super resolution models,a MRAN model based on multi-level attention mechanism is proposed.By adding the channel attention mechanism,the model can adjust the characteristic response of each channel adaptively and enhance the texture detail feature of pulmonary organ image.Using multilevel residual learning to gradually simplify the information flow and accelerate the network training;A multi-scale fusion module is added to obtain more complete structure information of the image.The Bicubic reconstruction branch was added to compensate for the feature loss during feature extraction.It shows that the MRAN model can obtain higher PSNR and SSIM values and achieve a better image reconstruction effects through lots of experiments.(2)To solve the problem of feature drift and the difficulty of using shallow features in Unet model,an improved Unet model based on quasi-pyramid is proposed,which is based on the Unet model.In this modle,the BN layer is added to accelerate the convergence of the network.The pyramid-like structure is used instead of the convolutional layer to obtain the features of each layer for the input of the subsequent layers,and the cross-layer fusion is used to realize the multiplexing of the feature information of each layer to increase the efficiency of feature transfer.The experiment shows that the segmentation effect of the improved Unet model is greatly improved compared with the original Unet,and the evaluation index value is superior to a similar algorithm.The correlation between the number of training data and the model performance is also explored through launching experiments.(3)To solve the problems of unbalanced training data and low precision of CHEXNET model,an improved CheXnet model is proposed.Dens Net121 is used as the basic architecture to complete the model construction.Various data augmentation methods are used to amplify the training set to solve the problem of unbalanced training data.CAM is used to focus on the lesion area and realize the visualization of the lesion area.Through comprehensive and systematic experimental verification,the average AUROC of CheXnet model with improved training strategies for intelligent diagnosis of 14 lung diseases reached 85.05%,which achieved more accurate intelligent diagnosis than similar algorithms.To sum up,the research results of this paper,on the one hand,can provide the reliable basis and reference for doctors' disease diagnosis,and on the other hand,provide a strong theoretical basis for intelligent medical treatment.