|
When an investigator is developing a new image-processing technique or manipulating an image-acquisition technique they are confronted with the question of whether the new technique will improve clinical diagnosis. A first approach is to look at individual physical properties of the image such as image contrast and resolution. Although these properties might be useful, it has long been known that the noise characteristics of the image system need to be taken into consideration to appropriately evaluate the quality of an image whether it will be used to detect, classify, and/or estimate a signal (Cunningham and Shaw, 1999). One useful measure of the noise characteristics is the noise-equivalent quanta (NEQ) that expresses the image noise in terms of the number of Poisson-distributed input photons per unit area at each spatial frequency (Wagner and Brown, 1985). The NEQ can be thought of as a measure inversely related to the amount of noise as a function of spatial frequency.
However, when the diagnostic decision involves a human observer, medical image quality can be defined in terms of human performance in visual tasks that are relevant to clinical diagnosis (Barrett, 1993). In this context, receiver operating characteristics (ROC) studies are the standard method of evaluating the impact of a particular image manipulation on clinical diagnosis. In these studies, the physicians scrutinize a set of medical images (under the different image-acquisition or processing conditions) and rate their confidence about the presence of the lesion. The investigator infers from these ratings a measure of performance known as the area under the ROC curve.
Often, the number of possible conditions is large and ROC studies become time consuming and costly because they require a large number of human observations. Other times, the investigator might want to optimize a parameter or a set of parameters. In such cases, the number of conditions suffers a combinatorial explosion, and therefore ROC studies become unfeasible. Thus it is desirable to develop a metric of image quality that could be used for fast evaluation and optimization of image quality but also would have the predictive power of ROC studies.
Computer-model observers are algorithms that attempt to predict human visual performance in noisy images and might represent the desired metric of image quality when the diagnostic decision involves a human observer and a visual task. Development of models to predict human visual signal detection in noise goes back to work by Rose (1948) who studied the detectability of a flat-topped disk embedded in white noise (see Burgess, 1999a, for a review). In the last two decades, many studies have concentrated on finding a model observer that can predict human performance across many types of synthetic backgrounds. More recently, model observers have been applied to real medical-image backgrounds. The hope is that eventually model observers will become common metrics of task-based image quality for evaluation of medical-image quality as well as optimization of imaging systems.
|
