The accuracy of radiographic detection of neurosurgical sponges was measured experimentally. Commercially available sponges were inserted into a cadaver in several paravertebral locations and radiographed. A total of 72 radiographs (90 sponges) were assessed, as well as an equal number of radiographs without sponges. The images were viewed by eight observers, whose responses showed marked variation in both false negatives and false positives; however, the two types of errors were compensatory and all responses were located along the same receiver operating characteristic curve. Mean false-negative and false-positive rates were both about 10%. In addition, false-negative rates varied between 3 and 25% according to the type of sponge.