Many researchers have tested and applied human-model observers as part of their evaluations of reconstruction methods for SPECT perfusion imaging. However, these model observers have generally been limited to signal-known- exactly (SKE) detection tasks. Our objective is to formulate and test scanning model observers that emulate humans in detection-localization tasks involving perfusion defects. Herein, we compare several models based on the channelized nonprewhitening (CNPW) observer. Simulated Tc-99m images of the heart with and without defects were created using a mathematical anthropomorphic phantom. Reconstructions were performed with an iterative algorithm and postsmoothed with a 3D Gaussian filter. Human and model-observer studies were conducted to assess the optimal number of iterations and the smoothing level of the filter. The human-observer study was a multiple-alternative forced-choice (MAFC) study with five defects. The CNPW observer performed the MAFC study, but also performed an SKE-but-variable (SKEV) study and a localization ROC (LROC) study. A separate LROC study applied an observer based on models of human search in mammograms. The amount of prior knowledge about the possible defects differed for these four model-observer studies. The trend was towards improved agreement with the human observers as prior knowledge decreased.

Date Published: 12 March 2009
PDF: 8 pages
Proc. SPIE 7263, Medical Imaging 2009: Image Perception, Observer Performance, and Technology Assessment, 72630R (12 March 2009); doi: 10.1117/12.813906

Published in SPIE Proceedings Vol. 7263:
Medical Imaging 2009: Image Perception, Observer Performance, and Technology Assessment
Berkman Sahiner; David J. Manning, Editor(s)