Research On Automated Detection Of Pulmonary Nodules Based On Chest Radiographs

At present,lung cancer is the second most common cancer.The 5-year survival rate of lung cancer patients is only 15%.If lung cancer can be detected early,the survival rate of lung cancer patients will increase from 15%to 49%.Early diagnosis of lung cancer is of great significance in improving the survival rate of lung cancer patients.At present,there are many kinds of imaging diagnosis methods for pulmonary disease diagnosis,in which X-ray radiograph has the advantages of low price,routine examination and low radiation.So x-ray chest radiograph is an important means of early detection and diagnosis of lung cancer.Early lung cancer in medical imaging is usually manifested as isolated pulmonary nodules.It is very important to detect pulmonary nodules in chest radiographs of early lung cancer patients and to take early treatment for patients.That there are overlapping anatomies in the chest radiograph is a challenge for radiologists to diagnose the disease.Computer-aided diagnostic method is applied to the image diagnosis technology,and it originated with the development of computer technology.Studies have shown that the use of computer-aided diagnosis can improve efficiency of radiologists reading and understanding chest radiographs.Detection of pulmonary nodules via computer-aided diagnosis can reduce misdiagnosis of doctors.The study of pulmonary nodules detection based on chest radiographs is gaining further attention from researchers.This study mainly focuses on pulmonary nodules detection in X-ray chest radiographs.This method is carried out from three steps,namely,lung fileds segmentation,pulmonary nodules candidate detection and nodules candidate classification.The difference between the method and previous studies is that the features set is not selected and deep learning model is used to classify candidates directly.The main contents of this paper are as follows:(1)Lung fileds segmentation is a mandatory step before image analysis of chest radiographs.The segmentation process of lung fileds is first pre-processed by luminance space normalization,and then the multi-resolution active shape model is used to segment.Finally the exact segmentation results are obtained.(2)According to the characteristics of pulmonary nodules,this paper presents a multi-scale weight convergence index filter improved method for pulmonary nodules candidate detection.Compared with the original algorithm,the improved algorithm is superior in terms of three performance indexs.(3)This paper applies AlexNet model of deep learning models to classify lung nodules candidate.According to characteristics of JSRT database is small and inconspicuousity of pulmonary nodules,the training data is augmented and processed by background trend correction.After nodules candidate detection,lung nodules candidate patches trunctated from chest radiographs are input to the trained AlexNet model.Finally the paper gets pulmonary nodules detection result images.Experiment shows that the algorithm solves target problem and achieves a good result.Correct rate of pulmonary nodules detection is 79.26%,and the average number of false positive per image is 5.6.The overall framework of this method reduces the process of pulmonary nodule segmentation and screening of pulmonary nodules feature set,reducing the overall complexity of the algorithm.