Misconceptions, confusion and politics surrounding CAD design and evaluation
CAD generates FROC data. Most researchers in this field are familiar with ROC data, which generates one rating per image and the location of suspicious region(s) is not considered in the analysis. Unlike ROC, FROC generates mark-rating data, where the mark is the location of a suspicious region and the rating is the confidence level that it is a true lesion. The number of mark-rating pairs per image is a-priori unknown, thereby making for a much more complex data structure that scares some "experts".
Because it does not give credit for unmarked non-diseased cases, which represent correct decisions, the FROC curve, proposed in 1969, is a poor descriptor of performance. Almost all CAD work to date has used the FROC curve. The correct procedure is discussed in detail in an upcoming book. The linked summary explains why the FROC curve is a poor descriptor of performance.