Objective assessment of low contrast detectability for real CT phantom and in simulated images using a model observer

The variability in doses and image quality used by different manufacturers of scanner models to reach similar diagnostic tasks has been proved to be wide. Image quality is frequently assessed performing human observer studies scoring the visibility of objects on CT images. These studies may become time consuming and expensive due to the high number of observers and observations required. Besides a bias can appear as the objects are arranged in patterns the observer knows beforehand. Besides, a great inter and intra-observer variability may exist. Computer model observers attempt to objectively predict human performance on the images and seem useful in investigating the influence of acquisition and reconstruction parameters and object size or shape on CT images. We have developed an objective statistical method with a model observer (non-prewhitening matched filter with an eye filter, NPWE) for detection tasks on CT, implementing characteristics of the selected kernel in the method. Images of the Catphan low contrast module (containing low contrast objects distributions) were acquired under different dose settings. Detectability (d') and proportion correct (PC) values were obtained for each object in the phantom. The results showed that d' increased with object size and mAs, and higher values were obtained as object contrast increased. Psychometric fits were performed and a visibility threshold of PC≥75% was established. In this way, the smallest visible object for each condition was obtained. To validate the model it was also applied to detection tasks on simulated white noise background images.