Study On Risk Factors Of Lung Nodules And Transfer Learning Assisted Diagnosis Of Chest Radiograph Based On Community Population

ObjectiveLung cancer has become a huge burden for human health.Population-based community screening is the best way for early diagnosis and treatment of lung cancer.Pulmonary nodules are the most common manifestation of early lung cancer.Screening pulmonary nodules is an important problem in the prevention and treatment of lung cancer.Research on the risk factors of pulmonary nodules is limited.Therefore,we conducted this study to fully understand the prevalence and influencing factors of pulmonary nodules detected by Low-dose computed tomography(LDCT),and to provide clues for the prevention and treatment of lung cancer.LDCT examination is currently an internationally recognized method for lung cancer screening,but it cannot be widely performed in China due to limitations of equipment and high cost.Chest radiographs is still the main method for lung cancer screening in most Chinese communities,while the diagnostic accuracy of chest radiographs restricts the screening efficiency.Artificial intelligence(AI)technology can improve the recognition of medical images,and thus provide new ideas for pulmonary nodule discrimination and lung cancer diagnosis.Based on the image database of the National Clinical Research Center and the screening population of the Tianjin Cancer Prevention and Demonstration Zone,this study will carry out a chest radiographic diagnosis research with artificial intelligence as the core technology.A new intelligent method for early diagnosis of lung cancer basing on chest radiographs will be developed after this study,which will then be applied in future community screening of lung cancer in China.MethodsA total of 6,579 lung cancer screening residents in Tianjin Cancer Prevention and Demonstration Zone were selected.The general demographic characteristics,smoking,family history of cancer,basic disease history,living habits,diets,indoor and outdoor air pollution were investigated.Chi-square test was used to compare the distribution of epidemiological factors between the two groups.Univariate and multivariate logistic regression were used to analyze the association between epidemiological factors and pulmonary nodules.The statistical indicators included odds ratio(OR)and 95%confidence interval(CI).Transfer learning(TL)was used to extract features from chest radiographs,and nonparametric tests were used to compare the distribution of 128 chest radiographs between the two groups.Stepwise logistic regression was used to establish the diagnostic model of pulmonary nodules.In the validation dataset,we used LDCT as the standard to evaluate the effectiveness of the model,including accuracy,sensitivity,and specificity.Receiver operating characteristic curve(ROC)was used to demonstrate the diagnostic ability of the deep learning model to distinguish between pulmonary nodules and non-nodules.Results1.Atotal of 6,579 residents were included,among which 1225 cases were detected with pulmonary nodules(18.6%).The majority of pulmonary nodules were micronodules and military nodules,435 cases(6.6%).Solid micronodules(≥5mm;<8mm)were followed by 272 cases(4.1%).Pulmonary nodules were most common in the upper lobe of the right lung(319 cases,26.0%).The lower lobe of the right lung and the lower lobe of the left lung were the next two(23.4%and 20.8%,respectively).2.There were statistically significant differences in gender,age,employment status and income between the two groups(P<0.05).Current or former smokers were associated with an increased risk of pulmonary nodules compared with never smokers.Passive smoking increased the risk of pulmonary nodules.Household coal burning was associated with an increased risk of pulmonary nodules.Regular exercise was associated with a reduced risk of pulmonary nodules in males.In addition,we found that drinking coffee at least 3 times a week,acidic diet and frequent use of air purification equipment are related to the occurrence of lung nodules.3.The training set consisted of 4956 images of 3049 individuals,including 1,514 patients with pulmonary nodules and 1,535 controls.Transfer learning was used to extract 128 features from clinical chest radiography images.Except for x0,x49,x75,x103,x117 and x119,other features between the two groups were statistically significant(P<0.05).The AUC,accuracy,sensitivity and specificity of the model were 81.5%,73.3%,72.0%and 74.7%respectively.4.The validation set included 2403 residents(3811 images),of whom 769 cases(32.0%)had at least one pulmonary nodule.The majority of pulmonary nodules were micronodules and military nodules,292 cases(12.2%).Pulmonary nodules were most common in the upper lobe of the right lung(201 cases,26.1%).Taking LDCT results as the standard,the model was validated in 2403 lung cancer screening residents.The AUC,accuracy,sensitivity and specificity of the model were 80.2%,78.8%,69.9%and 84.3%respectively.ConclusionIn this study,we analyzed the influencing factors of pulmonary nodules in lung cancer screening residents who underwent LDCT examination.Older age,smoking,passive smoking and household coal burning may increase the risk of pulmonary nodules.Females and exercise may reduce the risk of pulmonary nodules.Therefore,we recommend smoking cessation,avoiding passive smoking,reducing exposure to coal,and regular physical exercise to develop a healthy lifestyle and enhance the body’s ability to resist harmful exposure,thus reducing the incidence of pulmonary nodules.Transfer learning was used to extract features of chest radiographs and construct a model of pulmonary nodules.The effectiveness of the model was then evaluated in a community screening population of chest radiograph data.The model has good performance in the training set and validation set for the identification of pulmonary nodules.In the later stage,it can be popularized and applied to the future community screening of lung cancer in China,and improve the accuracy of screening technology and reduce the workload of radiologists.