Impact of anatomical noise on model observers for prostate SPECT

Scanning observers have been proposed for detection-localization tasks in medical imaging, but handling anatomical noise with these observers can be challenging. We have introduced visual-search (VS) observers as an alternative. The VS observer is a two-step process which mimics human perception through an initial search before a more detailed candidate analysis. Both the scanning and VS observers often outperform humans. In this paper, we analyze the performance of a VS observer with various inefficiency models as a means of matching human-observer performance. The observers were applied to prostate SPECT (single photon emission computed tomography) images. Prostate imaging protocols with In-111 and medium energy, parallel hole (MEPH) collimators were simulated. This was followed by a detailed localization receiver operator characteristic (LROC) study with human and model observers. Area under the LROC curve was used for observer performance evaluation. Results indicate that the VS observer applied with a combination of perceptual inefficiencies can quantitatively match human performance.