I121] Model observers applied to quality control in computed tomography

Abstract

Model observers are mathematical algorithms that can include different aspects of the human visual system and aim to reproduce human performance for simple detection tasks. Their use has extended in different medical imaging modalities, such as mammography, computed tomography (CT) and nuclear medicine among others, in the past few years. Several model observers have shown good correlation with human observers for simple detection and discrimination tasks in image quality phantoms with uniform backgrounds. Some model observers have been used by CT manufacturers to obtain clearance for their image quality claims, in particular for low contrast detectability (LCD), in the USA. LCD is frequently assessed in CT by human observers scoring the visibility of simple objects (such as disks) of different contrast and size in phantom images. These studies might be biased as, in general, the observer knows the distribution of the objects in the phantom beforehand. A great inter- and intra-observer variability might also appear. The models that will be studied in the course are: the non-prewhitening matched filter with an eye filter (NPWE) and the channelized Hotelling observer (CHO). The NPWE model includes a mathematical implementation of the contrast sensitivity function (CSF) of the human eye. There exist different published functions for the CSF based on human observer studies. The CHO model includes channels that can have different shapes (Gabor, Laguerre-Gauss, difference of gaussians…) depending on the target object; the channels aim to mimic the neurons triggering in the human visual cortex. Different aspects the implementation of model observers in image quality assessment, in particular in CT will be explored in this refresher course. Among them, how to choose the task, how to implement a model observer and its limitations, and also a short introduction on how to validate the model results with human observer studies. Practical examples about the use of these models in routine quality control, for example for low contrast detectability assessment, will be shown. Tips about how to start implementing model observers, in particular with freeware (imageJ), will be shared.