Lung nodules detection in chest radiography: image components analysis

We aimed to evaluate the effect of different components of chest image on performances of both human observer and channelized Fisher-Hotelling model (CFH) in nodule detection task. Irrelevant and relevant components were separated from clinical chest radiography by employing Principal Component Analysis (PCA) methods. Human observer performance was evaluated in two-alternative forced-choice (2AFC) on original clinical images and anatomical structure only images obtained by PCA methods. Channelized Fisher-Hotelling model with Laguerre-Gauss basis function was evaluated to predict human performance. We show that relevant component is the primary factor influencing on nodule detection in chest radiography. There is obvious difference of detectability between human observer and CFH model for nodule detection in images only containing anatomical structure. CFH model should be used more carefully.