Predicting the Benefit of Using CADe in Screening Mammography

Abstract

The goal of computer-aided detection (CADe) in screening mammography is to help radiologist avoid missing breast cancer. Thus when designing a CADe system it is important to know how different methods and parameters would affect radiologists’ ability to use the system effectively. Short of conducting an observer study or a clinical trial, this is not possible. In this paper, we present preliminary results on a model that can predict how many additional cancers a radiologist would detect, if they used a CADe system. The model uses the results of radiologists’ reading of a set of screening mammograms without using CADe to predict the probability that a radiologist would miss a cancer when reading without CADe and the probability that the radiologist would recall the woman if CADe flagged the missed cancer. In our initial study, 8 radiologists read 300 screening mammograms containing 69 cancers with and without CADe. Our model predicted that on average a radiologist would detect 4.7 extra cancers while the actual number of extra cancers detected per radiologist was 3.6. Bootstrapping across readers, the 95% CI for the difference between the predicted and actual number of extra cancers was [-1.52, 3.30]. The overall ability of the model to discriminate between lesions detected as a result of CAD flag and other lesion examinations was moderately high, with c-index of 0.77 (95% CI of [0.6, 0.91]). We are currently conducting a study with larger number of radiologists and cases to obtain better estimates of the accuracy of our model.