The aim of this study was to evaluate the accuracy of approximal caries diagnosis from digitized radiographs and digitally modified radiographic images, compared with conventional radiography. Twenty bitewing radiographs were digitized and from the digitized radiographs mirror-images were produced, resulting in 40 digital images. The caries progress at three approximal surfaces was graded by 12 dentists from conventional radiographs, digital unmodified images and digitally modified images. The image modifications were derived from the cumulative probability of grey values in the original digital image using one of five probability distributions: uniform, exponential, Rayleigh, hyperbolic cube root and hyperbolic logarithmic. Interobserver agreement was substantial. The unmodified digital images produced sensitivities comparable to radiographs but their specificities were lower. When the diagnostic task was to discriminate between `caries¿ and `no caries¿, the exponential, hyperbolic cube root and hyperbolic logarithmic modifications and radiographs performed equally well. With the decision cut off between `dentinal caries¿ and `no dentinal caries¿, the sensitivity of the hyperbolic logarithmic type of modification was statistically significantly superior to that of radiographs, but this modification was also associated with a statistically significant reduction of specificity. It is concluded that in particular the hyperbolic logarithmic modification can be an alternative to conventional radiography in incipient approximal caries diagnosis and restorative decision making.