Numerical observer for objective assessment on carotid plaque using spectral CT

We developed a novel numerical observer with signal variability to assess feature discrimination and compare the classification performance of carotid plaque using multi-energy, dual-energy, and conventional CT images. Our numerical observer was performed in two processing steps. First, a preprocessing step accounted for each spectral image bin using Channelized Hotelling Observer with a set of Laguerre-Gaussian channels, localized prewhitening, and localized matched filtering. The intermediate score of test statistics was obtained for each spectral image bin. The scores of all the CHOs were formed as a vector. Second, in an integration step, the test statistic vectors were used to make the classification decision of plaque features using the one-dimensional Hotelling Observer with additional prewhitening and matched filtering, yielding the figure-of-merit computation. To mimic a scenario in the clinical classification task, the known statistical distribution of signal variability was included. The validation of distinguishing two different plaque features in a known background was performed on simulated images using a digital anthropomorphic phantom with known compositions of inserted carotid plaques. The comparison of the performance for the calcified-plaque and fatty-mixed-plaque identification task showed that multi-energy CT outperformed dual-energy and conventional CT systems by 153.4%-174.3% (all p <; 0.01). For the differentiation task of the calcified plaque in the presence of iodinated blood, multi-energy CT yielded superior performance to the other CT systems by 158.1%-173.8% (all p <; 0.01). This proposed numerical observer with signal variability is an appropriate approach for the classification performance and can be extended to other clinical tasks such as kidney or urinary stone identification applications.