Image quality, the ideal observer, and human performance of radiologic decision tasks

Abstract

The quality of medical images must be quantified with reference to specific diagnostic tasks. Image quality is limited by fundamental physics, engineering limitations, radiation safety concerns, and imaging time constraints (among other things). There is now a gold standard for assessing human visual decision performance: the ideal Bayesian observer. Unfortunately, there are no mathematical tools to use this gold standard for realistically complex tasks. As an alternative, one can use the optimum linear discriminator (Fisher-Hotelling) model as a silver standard while en route to clinical realism. The goal of scientists working in the area is to develop mathematical models of human observers that will help equipment designers to optimize design trade-offs for specific diagnostic tasks. The current strategy is to modify the Fisher-Hotelling model to include certain limitations of the human observer visual system. The model must be both robust enough and mathematically tractable enough to be used to predict performance for clinical classification and estimation tasks. Statistical models also must be developed that describe realistic signals (lesions and abnormalities) and the normal patient structure that is the background in which these signals must be detected or identified.