Detection And Diagnosis Of Benign And Malignant Pulmonary Nodules In Chest CT Images Based On CNN Of Dual Shot Multi-boxes

The early manifestation of lung cancer is usually pulmonary nodules.The diameter of pulmonary nodules is mostly less than 30 mm.The detection of pulmonary nodules is a small target detection,which has some problems such as less information,low resolution and blurred image.Lung CT images are often used for lung cancer screening.With the continuous development of machine learning,deep learning and other technologies,the application of computer for automatic detection of lung nodules will reduce the workload of doctors and improve the recognition rate of lung nodules.Compared with other types of computer vision image recognition,there are still many difficulties in the detection of lung nodules.（1）The CT image of the lung is a three-dimensional structure,so it is difficult to calculate the number of lung nodules It costs a lot of resources;（2）Many nodule labeling is not accurate by medical experts,which leads to training errors.（3）At present,many algorithms have the problems of low recognition rate and high false positive rate,so the detection of pulmonary nodules is a problem worthy of study.In this paper,aiming at the problems of high false positive rate and low speed of detection algorithm in pulmonary nodule detection,a deep learning network model of double shot multi frame detection and recognition is proposed to assist doctors in the initial screening and diagnosis of pulmonary nodule.The algorithm mainly includes CT image preprocessing,matching coding calculation of target anchor frame and generated anchor frame,model training,pulmonary nodule recognition and model recognition,There are four parts in the evaluation.（1）In the preprocessing,the gray value of CT image is transformed into Hu（Hounsfield unit）value,the data set is expanded,and the world coordinates of CT image are transformed into image coordinates,so as to prevent other lung tissues from interfering with the detection of nodules.The morphology and threshold of pulmonary nodules are processed,and the mask is generated to separate the pulmonary interstitium from the pulmonary CT image.（2）In the matching coding of the target anchor frame and the generated anchor frame,the intersection ratio is calculated by designing anchor frame and target anchor frame with different proportions and sizes,and the label is marked.After coding by the coding function,the CT slices with nodules are loaded into the training model.（3）This model is based on SSD（single shot multi box detector）model,adding data enhancement processing,using bijection detector and progressive loss function calculation,more fully extract and learn small target features,more accurate detection,solve the problem of high false positive rate.In this paper,a large number of test experiments are carried out for the network model.Based on the open source data set LIDC / IDRI and Tianchi open source test data set,200 CT image test sets with pulmonary nodules are selected for training test evaluation.Most of the pulmonary nodules are 5mm ～ 10 mm in diameter.Different feature extraction networks and current popular target detection networks are used to test.The precision of resnet101 feature extraction network model is 86.4%,and the detection speed is fast.The data comparison shows that the algorithm has high recognition rate and sensitivity.