Semi-automated scoring of pulmonary emphysema from X-ray CT: Trainee reproducibility and accuracy

Objective We developed a semi-automated tool to quantify emphysema from thoracic X-ray multi-detector (64-slice) computed tomography (CT) for training purposes and multi-reader studies. Materials and Methods Thoracic X-ray CT was acquired in 93 ex-smokers, who were evaluated by six trainees with little or no expertise (trainees) and a single experienced thoracic radiologist (expert). A graphic user interface (GUI) was developed for emphysema quantification based on the percentile of lung where a score of 0 = no abnormalities, 1 = 1-25%, 2 = 26-50%, 3 = 51-75% and 4 = 76-100% for each lung side/slice. Trainees blinded to subject characteristics scored randomized images twice; accuracy was determined by comparison to expert scores, density histogram 15th percentile (HU ₁₅), relative area at -950 HU (RA₉₅₀), low attenuation clusters at -950 HU (LAC₉₅₀), -856 HU (LAC₈₅₆) and the diffusing capacity for carbon monoxide (DL_CO%pred). Intra- and inter-observer reproducibility was evaluated using coefficients-of-variation (COV), intra-class (ICC) and Pearson correlations. Results Trainee-expert correlations were significant (r = 0.85-0.97, p < 0.0001) and a significant trainee bias (0.15 ± 0.22) was observed. Emphysema score was correlated with RA₉₅₀ (r = 0.88, p < 0.0001), HU₁₅ (r = -0.77, p < 0.0001), LAC₉₅₀ (r = 0.76, p < 0.0001), LAC₈₅₆ (r = 0.74, p = 0.0001) and DL_CO%pred (r = -0.71, p < 0.0001). Intra-observer reproducibility (COV = 4-27%; ICC = 0.75-0.94) was moderate to high for trainees; intra- and inter-observer COV were negatively and non-linearly correlated with emphysema score. Conclusion We developed a GUI for rapid and interactive emphysema scoring that allows for comparison of multiple readers with clinical and radiological standards.