Improved detection of air trapping on expiratory computed tomography using deep learning

• DOI: 10.1371/journal.pone.0248902
• Publication Date: 2021-03-24T17:34:15Z
• Authors: Sundaresh Ram; Benjamin A. Hoff; Alexander J. Bell; Stefanie Galban; Aleksa B. Fortuna; Oliver Weinheimer; Mark O. Wielpütz; Terry E. Robinson; Beverley Newman; Dharshan Vummidi; Aamer Chughtai; Ella A. Kazerooni; Timothy D. Johnson; MeiLan K. Han; Charles R. Hatt; Craig J. Galban
• Keywords: Air trapping

Background Radiologic evidence of air trapping (AT) on expiratory computed tomography (CT) scans is associated with early pulmonary dysfunction in patients cystic fibrosis (CF). However, standard techniques for quantitative assessment AT are highly variable, resulting limited efficacy monitoring disease progression. Objective To investigate the effectiveness a convolutional neural network (CNN) model quantifying and AT, to compare it other measures obtained from threshold-based techniques. Materials methods Paired volumetric whole lung inspiratory CT were at four time points (0, 3, 12 24 months) 36 subjects mild CF disease. A densely connected CNN (DN) was trained using segmentation maps generated personalized method (PTM). Quantitative (QAT) values, presented as relative volume over lungs, DN approach compared QAT values PTM method. Radiographic assessment, spirometric measures, clinical scores correlated linear mixed effects model. Results found increase 8.65% ± 1.38% 21.38% 1.82%, respectively, two-year period. Comparison results intensity-based demonstrated systematic drop Dice coefficient (decreased 0.86 0.03 0.45 0.04). The trends observed consistent bronchiectasis, mucus plugging. In addition, be less susceptible variations deflation levels than approach. Conclusion effectively delineated scans, which provides an automated objective assessing patients.