A performance comparison of anthropomorphic model observers for breast cone beam CT images: A single-slice and multislice study.

PURPOSE: In this work, we investigate single-slice and multislice model observers which can predict human observer performance for simulated single-slice and multislice breast cone beam computed tomography (CBCT) images with a constant internal noise level.
METHODS: Breast background is generated based on a power spectrum of mammograms, and breast mass is modeled by a spherical signal. Human observer performance is evaluated for detecting 1 and 2 mm signals in different noise structures stemming from different reconstruction filters and image planes in a Feldkamp-Davis-Kress reconstruction. To predict human observer performance, we use single-slice channelized Hotelling observer (i.e., ssCHO) and multislice CHO (i.e., msCHOa and msCHOb) with dense difference-of-Gaussian and Gabor channels. In addition, we use single-slice nonprewhitening observer with an eye-filter (i.e., ssNPWE) and multislice NPWE (i.e., msNPWEa and msNPWEb), where ms-a model estimates the template for each image slice and ms-b model estimates the template for the central slice. For NPWE, we use the most common eye-filter with a peak value at a frequency of 4 cyc/deg. In addition, we propose an eye-filter with a peak value at a frequency of 7 cyc/deg which shows good correlation with human observer performance in single-slice breast CBCT images. Channel and decision variable internal noise are used for CHO, and decision variable internal noise is used for NPWE. The internal noise level is determined by comparing human and model observer performance for single-slice images, after which the same level is used for the multislice model observers.
RESULTS: For single-slice images, all model observers predict human observer performance well. When the same internal noise level for the single-slice model observer is used for the multislice model observer, CHO with channel internal noise produces a higher performance than the human observer. In contrast, msCHO and msNPWEb with decision variable internal noise produce a similar performance to the human observer. Especially, ssNPWE and msNPWEb with the proposed eye-filter predict the human observer performance better than the other model observers for different noise structures.
CONCLUSIONS: ssCHO/ssNPWE and msCHO/msNPWEb with decision variable internal noise can predict human observer performance for single-slice and multislice images with the same internal noise level. In the presence of breast anatomical background, ssNPWE and msNPWEb with the proposed eye-filter predict human observer performance better than the other model observers for different noise structures.