Implications of unchanged detection criteria with CAD as second reader of mammograms

In this paper we address the use of computer-aided detection (CAD) systems as second readers in mammography. The approach is based on Bayesian decision theory and its implication for the choice of optimal operating points. The choice of a certain operating point along an ROC curve corresponds to a particular tradeoff between false positives and missed cancers. By minimizing a total risk function given this tradeoff, we determine optimal decision thresholds for the radiologist and CAD system when CAD is used as a second reader. We show that under very general circumstances, the performance of the sequential system is improved if the decision threshold of the latent human decision variable is increased compared to what it would have been in the absence of the CAD system. This means that an initial stricter decision criterion should be applied by the radiologist when CAD is used as a second reader than otherwise. First and foremost, the results in this paper should be interpreted qualitatively, but an attempt is made at quantifying the effect by tuning the model to a prospective study evaluating the use of CAD as a second reader. By making some necessary and plausible assumptions, we are able to estimate the effect of the resulting suboptimal operating point. In this study of 12 860 women, we estimate that a 15% reduction in callbacks for masses could have been achieved with only about a 1.5% relative decrease in sensitivity compared to that without using a stricter initial criterion by the radiologist. For microcalcifications the corresponding values are 7% and 0.2%.