An ROC evaluation of six systems for chest radiography.

A comparison of the diagnostic performance of six methods of performing posteroanterior radiography of the chest was carried out using images of an anthropomorphic phantom. Experienced observers were asked to detect nodular (coin) lesions inserted in the lung fields and mediastinal areas of the phantom and the results were subjected to an ROC analysis. Results showed that there were only marginal, non-significant differences in ROC performance (Az) between the individual systems: Lanex, Dupont Ultravision, Kodak HC and Kodak VHC and Fuji AC (P>0.1) but that a league table of different (Az) scores could be formed. The AMBER system was superior to the other systems at various levels of significance for mediastinal lesions (P0.1), for lung lesions (P0.1), and for lesions in any location (P01). The computed radiography system Fuji AC performed least well in this league table but operational factors were felt to have contributed to this outcome. Significantly the system was tested under the normal operating conditions for its clinical setting, that is at a lower kV than the other systems and without the use of a grid. Comments and suggestions are made on why computed radiography may underperform if (a) it is not operated at optimum acquisition conditions and (b) its unique post-processing attributes are not exploited