Deep neural networks based lung function prediction using CT images
The latest epidemiological survey reveals a high and escalating prevalence rate of chronic respira-tory diseases,such as Chronic obstructive pulmonary disease(COPD)and asthma,posing a significant public health threat.Computer tomography(CT)has emerged as a convenient and noninvasive method for evaluat-ing pulmonary function,however,in existing computer-aided lung function evaluation method,handcrafted methods are insufficient and existing neural network methods are less effective in extracting features from small datasets with high noise and sparse data.In this study,a lung function prediction network(LFP-ResNet)is introduced to predict lung function from CT images.Firstly,a multi-level contextual feature fusion(MCFF)method is proposed to extract diverse features that represent the pulmonary texture and morphology effectively.Secondly,a three-dimensional(3-D)residual network is used to guarantee the spatial heterogene-ity of CT image sufficiently.Finally,a dataset containing both healthy population and patients with chronic re-spiratory diseases are constructed and used to compare the proposed methods with other state-of-the-art meth-ods.The results demonstrate that the proposed MCFF strategies are more efficient in extracting features from a sparse matrix with noise than other feature extraction methods.Moreover,the constructed LFP-ResNet ex-hibits better predictive performance in the pulmonary function prediction task.
Computed tomographyDeep learningMulti-task learningPulmonary function testChronic obstructive pulmonary disease
国家科技期刊平台
NSTL
万方数据
