Image quality can be defined objectively in terms of the performance of some "observer" (either a human or a mathematical model) for some task of practical interest. If the end user of the image will be a human, model observers are used to predict the task performance of the human, as measured by psychophysical studies, and hence to serve as the basis for optimization of image quality. In this paper, we consider the task of detection of a weak signal in a noisy image. The mathematical observers considered include the ideal Bayesian, the nonprewhitening matched fiter, a model based on lineardiscriminant analysis and referred to as the Hotelling observer, and the Hotelling and Bayesian observers modified to account for the spatial-frequency-selective channels in the human visual system. The theory behind these observer models is briefly reviewed, and several psychophysical studies relating to the choice among them are summarized. Only the Hoteiling model with channels is mathematically tractable in all cases considered here and capable of accounting for all of these data. This model requires no adjustment of parameters to fit the data and is relatively insensitive to the details of the channel mechanism. We therefore suggest it as a useful model observer for the purpose of assessing and optimizing image quality with respect to simple detection tasks.Image quality, for scientific and medical purposes, can be defined in terms of how well desired information can be extracted from the image. In other words, image quality is measured by the performance of some "observer" on some specific task (1-3). The observer can be a human, such as a physician trying to make a diagnosis, or it can be a mathematical model or a computer algorithm. The tasks can be divided generically into classification and estimation tasks (4). In medical applications, an example of a classification task would be lesion detection, while an estimation task might be determination of the volume of blood expelled from the heart on each beat.For classification tasks performed by a human observer, psychophysical studies and receiver operating characteristic (ROC) analysis provide a reproducible, quantitative measure of image quality (2,3,5), but such studies are time consuming and require large numbers of images. Moreover, they do not provide an easy way to see how image quality is related to various parameters of the imaging system or processing algorithm. For these reasons, there is considerable interest, especially in the radiological literature (6-8), in mathematical model observers. If the ultimate observer will be a human rather than a machine, the objective of the model is to predict accurately the performance of the human. Then the model observer can be used for system evaluation and optimization with some assurance that the system that is best for the model is also best for a human. Model observers may also be usedThe publication costs of this article were defrayed in part by page charge payment. This article must therefore be hereby mark...