The ability of an observer to detect variations in size of a geometrical image feature have been investigated using receiver operating characteristic (ROC) analysis. Three types of image were constructed using computer graphics: disc-shaped targets of variable radius, model chest radiographs showing a variable heart diameter and model arterial angiograms with variable vessel width. Five factors were investigated: observer experience, variation of detectability with theoretical signal-to-noise ratio, the prior probability of the presence of an abnormality, viewing distance, and uncertainty in the location of an abnormality. In all but one experiment, excellent agreement was found between measured detectabilities and the predictions of signal detection theory, providing an initial practice session was included for each observer. No significant variation in detectability was found using six different prior probabilities and two different viewing distances, and the reduction in detectability for a four-alternative location task was in good agreement with theoretical predictions. The high statistical efficiencies found for the detection of geometrical signals suggest that the levels of observer "internal" noise arising from decision-making processes during an ROC experiment are very low.