Challenges for the evaluation of diagnostic imaging systems with nonlinear behavior
The adoption of iterative reconstruction algorithms (ongoing in the case of X-ray Computed Tomography (CT), and already well established in emission tomography modalities like SPECT, PET), and the use of additional imaging processing techniques, are resulting in imaging systems with nonlinear behavior. This alters the image noise properties, sometimes leading to noise patterns unfamiliar to the radiologists. The spatial correlations and the nonlinear behavior also challenge some fundamental assumptions used in assessing the image quality, making inoperable several more traditional metrics such as pixel variance, contrast to noise ratio (CNR), modulation transfer function (MTF), and related metrics expressed as signal to noise ratios (SNR). In this talk we will review some of these challenges and present alternative evaluation methods based on assessing the performance of a given task. One category of such task-based methods involves detection of small, low-contrast signals on noisy backgrounds. We can further distinguish between the particular case when the signal and the background are known exactly (SKE/BKE tasks), and the cases when the some signal features are not precisely known. One such case of wide clinical interest is the detection of signals at unknown locations. While the SKE/BKE case can be treated by extending the classical signal detection theory to the multiple dimensional case, the problem of unknown signal detection proves to be less tractable analytically, with only approximate solutions being proposed. Here we will discuss several practical approaches and applications to CT and PET image quality evaluation.