Intelligent Diagnosis Of Multiple Class Lung Diseases Based On Improved FasterRCNN

Affected by environment,climate and lifestyle,the incidence rate of lung disease in China is increasing year by year.Lung diseases not only affect the lungs,but also cause many complications,such as tracheitis,heart disease,lymphatic system disease,etc.CT image is the most effective non-invasive detection technology for lung diseases because of its advantages of thin layer,high definition and low noise.It is widely used in lung disease screening and auxiliary diagnosis.With the development of science and technology,machine-readable film is used to assist doctors to screen the location of lesions and diagnose diseases.The traditional machine reading film is usually obtained by experienced doctors using medical knowledge to observe the intuitive morphological characteristics of lesions,which is subjective.If the machine can automatically extract the features learned from CT images and find out the CT images with lung diseases,the results will be more objective and accurate.In recent years,deep learning has been widely used in image classification,speech recognition and target detection,and many large hospitals have also used deep learning to assist doctors in the field of medical diagnosis.At present,the research of lung CT image focus screening based on deep learning mainly focuses on the diagnosis of pulmonary nodule,a single disease,and the research on the recognition and diagnosis of multiple lung diseases is very little.However,there are many kinds of lung diseases.There may be multiple lesions and diseases in the same patient’s lung.Repeated screening not only wastes time,but also has no reasonable use of resources.Therefore,this paper uses fast RCNN,a classical deep learning algorithm for target detection,and optimizes it to identify pulmonary nodules,cords,arteriosclerosis or calcification,lymph node calcification and other lung diseases.Through a screening of a medical record,we can get the location and category of all suspected lesions,without repeated screening,and improve the diagnosis efficiency.The data set of this paper is from Tianchi competition:global data intelligence competition(2019)-training set for intelligent diagnosis of multiple diseases of lung CT in "digital human body" competition.Firstly,the data are analyzed and preprocessed.In the part of descriptive analysis,we introduce the meaning of all kinds of information contained in the mhd file,the proportion of four kinds of lesions in the total,and the Hu value distribution of lung CT image.In the preprocessing part,firstly,the labeled coordinates are transformed into voxel coordinates to show the location of the focus in the slice image more clearly;secondly,the lung tissue is segmented,and the image segmentation method based on the floodfill algorithm is used.Firstly,the image is binarized,then the irrelevant background such as machine is removed,then the hole is filled based on the morphological method,and the floodfill algorithm is used to improve the outer contour of lung tissue,finally,is expanded to include the parts prone to arteriosclerosis or calcification and lymph node calcification;then,the data is standardized according to the window width of CT image;finally,the data set is divided into training set and validation set.In this paper,an improved Faster RCNN network is proposed for target detection of these four lung diseases.The original backbone network of Faster RCNN is changed to a deeper resnet plus FPN network.In addition,the ROI pooling layer is changed to ROI align,and bilinear interpolation is used to find the corresponding position of the preselected box on the original picture more accurately.This paper compares the target detection effect of the original Faster RCNN network,the improved Faster RCNN network with resnet50+FPN as the backbone network,and the improved Faster RCNN network with resnet101+FPN as the backbone network.The commonly used technical evaluation index of the target detection task mAP is used.Experiments show that the improved Faster RCNN with resnet101+FPN as the backbone network can more effectively diagnose a variety of lung diseases.