The "memory effect" for repeated radiologic observations

It is assumed that memory has a role to play in repeated radiologic observation studies. The main objective of this study was to examine this assumption and evaluate the effect that memory may have on receiver operating characteristic (ROC) methods.

A two-center observer study was performed with a total of 24 experienced radiologists. Over two viewings, chest radiographs showing the tip of a central line in either the superior vena cava or the azygos vein were presented. Half of the images were changed between the two viewings. The participants' attention was directed on the first reading to the position of the central line. At the second reading, the participants were asked to assign a confidence score on a 6-point scale about whether each image had been included in the first reading.

For the images that were scored as "definitely included" in the first viewing, readers at our two centers recalled only an average of 2.5 and 4.9 of the 20 repeated images, which is close to a random allocation of images to each score. As the confidence levels diminished for positive identification of repeated images, the numbers of correct answers increased. For images scored as not having been previously included, the numbers of correct answers remained low suggesting that identification of nonrepeated images is poor. Images with a greater number of incidental abnormalities and with more striking abnormalities were recognized more accurately than those with fewer and less striking abnormalities.

This study shows a "memory effect" when the same images are presented at a second viewing within a small interval period. This effect appears to occur mainly at low confidence levels. These results suggest that including images with obvious incidental abnormalities in reader performance studies should be avoided.