A comparison of three statistics for detecting differences in digitized dental radiographs: a simulation study

Because of methodology-induced structural differences in dental radiographs, determination of change has always depended upon expert interpretation. However, new methods should be able to considerably reduce structured error in digitized subtracted images. Once true change in density is obscured only by random variation in pixel density, statistical methods may be brought to bear on the problem of detecting change. The most appropriate statistic is not obvious, however, since density change can be quantified with respect to both magnitude and dimensional extent. Whereas mean density loss is often intuitively defined as the average density of those pixels losing density (to preclude gaining pixels from offsetting losing pixels), the extent of change may be defined in a variety of ways. In this study, extent was defined as either (a) the total number of pixels losing density, or (b) the size of the largest cluster of losing pixels. The object was to evaluate the comparative statistical power of three possible statistics (based on mean density, number of losing pixels, and size of largest losing cluster) for detecting change.

In a series of simulations of comparative clinical trials, density was reduced in the centre of 1600-pixel square regions of interest by either one or 10 grey-scale units, and t-tests, based on the three statistics, were then compared for their ability to detect differences.

Each of the three statistics was shown to exhibit superior relative power under particular conditions of loss magnitude, loss distribution, and pixel threshold for change.

Selection of the appropriate statistic for identifying change between radiographs will require further information about the anticipated distribution of density changes for the different disease processes under investigation.