Comparison of microscopy and radiography as gold standards in radiographic caries diagnosis.

OBJECTIVES

To compare the effect of the choice of gold standard on the diagnostic outcome of approximal caries detection in original and compressed digital radiographs.

METHODS

116 extracted teeth radiographed with a storage phosphor system constituted the original images. These images were compressed at 1:20 and 1:33 with the JPEG irreversible compression standard. Five radiologists scored the three sets of images for the presence of approximal caries on a five-rank confidence scale. The radiographic scores were validated by stereomicroscopy (the true gold standard). The individual ROC areas for the five observers were used to select the worst (obsworst) and the best (obsbest) performer: Their scores in the original images were used as the second and third 'gold standards' for the remaining observers. Mean ROC areas for the three observers with the three types of images were calculated using these two new 'gold standards'. Differences between the ROC areas when using microscopy, obsworst, and obsbest as the 'gold standard' were compared.

RESULTS

The mean ROC areas in the original images were 0.66, 0.74 and 0.91 using the true gold standard and obsbest and obsworst as the 'gold standards' respectively. The difference between the true gold standard and obsworst was statistically significant (P < 0.001). The mean ROC areas using the true gold standard decreased with increasing compression whereas they were constant or increased using obsworst and obsbest as 'gold standards', respectively.

CONCLUSIONS

Accuracy in approximal caries diagnosis was significantly higher when an observer was the 'gold standard' than when the true gold standard was obtained by microscopy. Paradoxically, the compressed, degraded images were more accurate than the originals when an observer was the 'gold standard' while they were less accurate with the true gold standard. Thus, results obtained using observers' scores from the radiographs which are being evaluated, as validation for the presence of caries may mislead the clinician.