Improved imaging diagnosis by sequentially combined confidence judgments.

Abstract

Consultation among radiologists improves diagnostic efficacy. However, it is unclear in practice when consultation is indicated, and consultation may be logistically difficult. We present a computerized method of noninteractive consultation that supersedes these problems, using radiograph readers' confidence judgments in their decisions. Individual categorical confidence judgments of excretory urogram diagnoses were transformed into diagnostic probabilities, then combined sequentially by computer simulation with other radiologists' probabilities until sufficient group confidence was achieved. Appropriate stopping rules were defined empirically. Analysis of cost effectiveness of this model--considering the cost of sequential interpretations and a range of costs for false-positive and false-negative errors--indicates that mathematically combined sequential decisions are more efficacious and less costly than individuals interpreting urograms alone.