Purpose: Channelized Hotelling observer (CHO) models have been previously implemented for x-ray angiography; however, they have only accounted for static test objects and uniform background. The purpose of this work was to develop theory and methods to apply statistical performance tests to x-ray angiography systems with spatial and temporal non-stationarity, including anatomical background and moving test objects.
Methods: Unprocessed images of static and moving iodine-contrast test objects with both uniform and anatomical (anthropomorphic phantom) backgrounds were acquired with a commercial x-ray angiography system. Uniform background images were used to determine task-specific efficient sets of CHO Gabor channels via a novel backward elimination method and to compare the performance of the efficient CHO to that of a quasi-ideal Hotelling observer (HO). The efficient CHO was then adapted to perform a single-sample multivariate statistical test to enable estimation of dâ€™ for moving test objects under anatomical background. Estimation and correction of bias due to finite sampling and system non-stationarity was performed for all experimental conditions.
Results: Channel reduction between 60-80% of the initial channel set (96 channels) was achieved. The efficient CHO related linearly with the quasi-ideal HO with an efficiency >90% across experimental conditions. dâ€™ of the single-sample CHO applied to moving test objects and uniform background was equivalent to the standard CHO, albeit with less precision. Under anatomical background, dâ€™ of the single-sample CHO highly reflected local attenuation properties and related linearly with dose per frame as expected from quantum SNR theory. The proposed dâ€™ bias correction strategy effectively mitigated the impact of bias from non-stationary noise on estimates of dâ€™, particularly for small objects and low doses.
Conclusion: This work demonstrates the theory and methods of the statistical CHO can be adapted to accommodate imaging systems with spatial and temporal non-stationarity. Bias management was critical to maintain model accuracy.