A noise mechanism for model observers

Model observers may be employed as surrogates for human observers for various parameter optimization tasks that involve evaluation of imaging systems and processing algorithms. For a lesion localization-detection task, model observers based on statistical detection theory typically demonstrates superior performance compared to human observers. Inclusion of an internal-noise-mechanism can allow a model observer to better match humans. In a sequential reading task, human-observer internal noise may depend in a complicated manner on the ensemble of image data in addition to the particular image being viewed in the sequence. There may be other data-independent components of internal-noise as well. In this work, we propose an internal-noise model that takes this conditional dependence on the data-ensemble into consideration. Different ways of defining the ensemble lead to different sets of parameters for the noise-model, which in turn lead to different dynamics for the parameter-optimization curves. We explore these various cases using the channelized non-prewhitening (CNPW) observer. Our dataset consisted of Ga-67 hybrid SPECT/CT data, which was reconstructed using various combinations of attenuation compensation (AC), scatter Compensation (SC) and resolution compensation (RC). We generate optimization curves for these different combinations and compare the results with the noise-less case.