Application Of Low-dose CT And Deep Convolutional Neural Networks For Chronic Obstructive Pulmonary Disease Classification And Bronchoscopy

Classification of chronic obstructive pulmonary disease on chest CT images using deep learning:a multicohort studyPurpose The pathogenesis of chronic obstructive pulmonary disease(COPD)depends on emphysema and small airway remodeling,which can be displayed by computed tomography(CT).We intend to use a deep convolutional neural network(CNN)to predict the presence of COPD based on CT images.Methods Subjects who underwent non-contrast chest CT and pulmonary function testing(PFT)within three days were retrospectively included in one hospital to train and validate a DenseNet 201 CNN.Subjects from another hospital served as external test set.Subjects were divided into COPD and normal group based on PFT results.Lung parenchymal and bronchial wall images were used for deep learning.Binary logistic regression was applied to combine the classification probability based on lung parenchyma and bronchial wall.The percentage of voxels below-950HU and the square root of wall area at 10-mm lumen perimeter were calculated as traditional CT quantification parameters.Results 272 cases(training and validation cohorts)and 144 cases(external test cohort)were included from hospital A and B respectively.In the external test cohort,the accuracy of CNN to classify COPD was 75.0%(95%CI:67.1-81.8%)and 79.2%(71.6-85.5%)based on lung parenchyma and bronchial wall,respectively.The logistic regression model improved the accuracy to 83.3%(76.2-89.0%),significantly higher than 64.6%(56.2-72.4%)for traditional quantification methods(p<0.0001).The accuracy of our deep CNN model to classify five-way GOLD stage was 79.2%and 69.4%in validation and external test cohort,respectively.Conclusion The CNN successfully classifies the presence of COPD based on CT images of lung parenchyma and bronchial wall,suggesting the potential to detect COPD based on extensively used chest CT.Airway quantification using adaptive statistical iterative reconstruction-V on wide-detector low-dose CT:a validation study on lung specimenPurpose To evaluate the accuracy of airway quantification of adaptive statistical iterative reconstruction veo(ASIR-V)on low-dose CT using a human lung specimen.Methods A lung specimen was scanned on Revolution CT with low-dose settings(20mAs,40mAs and 60mAs/100kV)and standard-dose setting(100mAs/120kV).CT images were reconstructed by using lung kernel with eleven ASIR-V levels from 0%to 100%with 10%interval.The left main bronchial wall and lumen were selected to measure the signal noise ratio(SNR)and noise.Wall area percentage(%WA)and wall thickness(WT)were measured.Results Radiation dose of 20mAs,40mAs and 60mAs low-dose settings reduced by 87.6%,75.2%and 62.8%compared to that on standard-dose respectively.Low-dose settings significantly decreased image SNR and increased noise.ASIR-V level exponentially improved image SNR and linearly decreased image noise.The mean airway measurement ratios of low-dose to standard-dose were within 2%variation for%WA and within 3%variation for WT.%WA and WT values showed no obvious correlation with ASIR-V levels.Conclusion ASIR-V improved image quality in low dose CT.However,low-dose settings and ASIR-V strength did not significantly influence airway quantification values,although variation of the measurements slightly increased as dose reduced.