Effect of observer instruction on ROC study of chest images.

摘要: Receiver-operating characteristic (ROC) analysis has been used in many medical imaging applications during the past decade. In order to ensure that reader-confidence ratings are analyzable (well distributed meet convergence requirements of curve-fitting algorithms) and meaningful (limit extrapolation data), investigators train readers specifically for this purpose. No experimental data available concerning possible effects such training on results ROC studies. We performed a multi-observer, multi-disease study which 300 chest images were rated by four radiologists before after they trained provide well-distributed confidence ratings. The indicate our set, reader disease-specific accuracy was not significantly affected process interstitial disease pneumothoraces. However, two detection nodules (P less than 0.05), overall one classification normal versus abnormal 0.01). Thus, spite difficulties associated with performance studies free-reading environment, should carefully consider any intervention prior conducting