Detection Of Pulmonary Nodules On Chest Radiography Based On Convolution Neural Network

As one of the common malignant tumors,lung cancer has a very high mortality rate,but if it can be detected early and treated promptly,it will greatly improve the survival rate of patients.Lung cancer often manifests as pulmonary nodules in the early stage.Thus the accurate detection and diagnosis of the pulmonary nodules are very important for the discovery and treatment of early stage lung cancer.As the most common medical imaging diagnosis method,chest radiograph has the advantages of small radiation dose and low cost,and is widely used in physical examination.The detection of pulmonary nodules is highly subjective,and it is difficult for clinicians to ensure that each chest piece is viewed with the same standards and concentration.Therefore,a computer-aided detection system with superior performance can reduce the workload of doctors and improve the diagnostic accuracy..Due to the two-dimensional imaging characteristics of X-rays,pulmonary nodules overlap with ribs and other organs in the chest radiograph,increasing the difficulty of detection.The current method for detecting chest and pulmonary nodules is difficult to achieve high sensitivity,low false positive rate and real-time.In view of the above situation,this thesis proposes a rapid and accurate method for detecting pulmonary nodules in the chest X-ray images.The main innovations and contributions include:(1)In the image preprocessing stage,this thesis proposes a U-net-based rib suppression residual network for the problem of leakage of pulmonary nodules caused by rib interference in chest radiograph.A ribless chest radiograph is used as a training label in a supervised learning manner.The network uses a U-shaped structure to ensure image detail through bounce connections and residual learning strategies.Experiments show that the chest radiograph image processed by the rib suppression network can eliminate the negative effects of rib image,highlight the nodule and lung texture features,and reduce the difficulty of lung nodule detection task.Under the same lung nodule detection model,the chest radiograph data processed by the rib suppression network increased the mean average precision of the lung nodule detection model by 2.5%.(2)In view of the small size of the lung nodules,this thesis uses the object detection network RetinaNet with feature pyramids structure as the basic model.Considering that the distribution of nodules in the chest radiograph has certain regular characteristics,this thesis proposes a lung nodule detection network LSFNet that integrates positional and scale information,and combines position and scale information with image features as the basis for judgment.In the public dataset JSRT,the sensitivity of this method is 93%with an average of 4.7 false positive results,89%for an average of 2 false positive results,and the highest sensitivity is close to the current highest level,but The method in this thesis has higher sensitivity in the case of a lower proportion of false positive results.(3)Integrating rib suppression and other pretreatment methods with the lung nodule detection network to achieve the chest X-ray nodule detection system.Ten chest radiographs with nodules were used to test the system.The sensitivity was 100%,and the average number of false positive nodules was 2.Meanwhile,the average detection time was 0.7 seconds.