Abstract
X-ray computed tomography (CT) has become widely used as a diagnostic tool in radiology and other medical specialties. CT images are noisy, and the dependence of noise on the radiation dose to the patient ensures that there will always be interest in how much noise can be tolerated before diagnostic performance is significantly degraded. However, relatively little is known about how human observers extract information from such images in the presence of textured noise and normal anatomical variability.
In this study, we evaluate human observers in relatively simple forced-localization studies across two target sizes and 12 different Gaussian noise textures. The noise textures have power-spectra that reflect different levels of anatomical variability, and different levels of processing (apodization) in the image reconstruction process. We analyze observer performance using efficiency with respect to the Ideal Observer as a benchmark for how effectively observers can extract diagnostic information and classification images to demonstrate how they extract it.
We find efficiency in these forced localization tasks is generally quite high, ranging from 20% to almost 80%. The lowest efficiency is found in conditions with the least “natural” statistics. The classification images show modest adjustment of bandpass frequency weights in response to the different noise textures, and also suggest that a substantial component of the lowest efficiency conditions comes from inability to effectively use the lowest spatial frequencies. We also show that the classification images, when used as a scanning linear template, can explain 86% of variance in average efficiency across all conditions.