Observer performance in assessing anemia on thoracic CT

The purpose of this study was to evaluate the ability of expert reviewers to differentiate an anemic from a nonanemic state on the basis of visual assessment of the relative attenuation of blood in the left ventricle on noncontrast thoracic CT images and to compare reviewer performance with quantitative measurement of CT density in Hounsfield units.

One hundred two noncontrast thoracic CT examinations were qualitatively reviewed by three independent reviewers. Hounsfield unit measurements of the blood in the left ventricle were recorded by a fourth individual. Anemia was defined as a hemoglobin level of less than 10 g/dL. Receiver operating characteristic (ROC) analyses of expert reviewers were compared with measured Hounsfield units.

Hounsfield unit measurements performed significantly better than subjective reviewer analyses for differentiation of an anemic from a nonanemic state (area under ROC curve = 0.85 vs 0.72, 0.70, and 0.69; 95% confidence interval, 0.78-0.92 vs 0.63-0.81, 0.61-0.79, and 0.60-0.78, respectively; p < 0.05). With use of a CT density threshold of 35 H, the sensitivity for anemia was 76% and specificity was 81%, whereas the sensitivity of three reviewers was 40-72% with a specificity of 60-83%. Interobserver agreement was found to be poor by kappa statistic (0.0906-0.2128). The correlation coefficient for the analysis of Hounsfield unit versus hemoglobin level was 0.72. Separating data by patient sex revealed a correlation coefficient of 0.81 for men versus 0.52 for women, although the actual regression lines were not statistically different (p > 0.05).

Despite expert reviewer analyses, subjective evaluations of blood attenuation characteristics are prone to inaccuracy and show poor interobserver agreement. Quantitative measurements of CT density in Hounsfield units should be performed to accurately differentiate an anemic from a nonanemic state when serum hemoglobin levels are not readily available.