A Multiclass Model Observer for Multislice-Multiview Images.

Abstract

A human-model observer for tumor detection-localization studies featuring multislice-multiview (or volumetric) image displays has been introduced. This volumetric observer, an extension of multiclass linear observers previously tested with single-slice and multislice displays, produces rating and localization data by integrating perception measurements from the different image views. A channelized NPW (CNPW) version of the observer was evaluated against humans for a background-known-exactly (BKE) detection task involving localization of Tc-99m Neotect lesions in simulated SPECT lung images. An LROC study evaluated two RBI reconstruction strategies that used different combinations of corrections for attenuation, scatter, and distance-dependent system resolution, and coronal, sagittal, and transverse slices were presented to the observers. Model-observer ranking of these strategies did not match that of the humans. Follow-up studies exploring several possible remedies for the model observer, including strategy-specific search regions and an internal-noise mechanism, showed little change. Future work will examine variations from the BKE assumption as a means of reconciling the rankings.