Computer-aided diagnosis of emphysema in COPD patients: Neural-network-based analysis of lung shape in digital chest radiographs

Abstract 

Several abnormalities of the shape of lung fields (depression and flattening of the diaphragmatic contours, increased retrosternal space) are indicative of emphysema and can be accurately imaged by digital chest radiography. In this work, we aimed at developing computational descriptors of the shape of the lung silhouette able to capture the alterations associated with emphysema. We analyzed two-sided digital chest radiographs from a sample of 160 patients with chronic obstructive pulmonary disease (COPD), 60 of which were affected by emphysema, and from 160 subjects with normal lung function. Two different description schemes were considered: a first one based on lung-silhouette curvature features, and a second one based on a minimal-polyline approximation of the lung shape. Both descriptors were employed to recognize alterations of the lung shape using classifiers based on multilayer neural networks of the feed-forward type.

Results indicate that pulmonary emphysema can be reliably diagnosed or excluded by using digital chest radiographs and a proper computational aid. Two-sided chest radiographs provide more accurate discrimination than single-view analysis. The minimal-polyline approximation provided significantly better results than those obtained from curvature-based features. Emphysema was detected, in the entire dataset, with an accuracy of about 90% (sensitivity 88%, specificity 90%) by using the minimal-polyline approximation.

Keywords: COPD, Emphysema, Digital radiography, Shape analysis, Neural networks