EM clustering for holistic search in human-model observers

Abstract

Mathematical models of human observers for detection tasks are of interest in medical imaging analysis, but current model observers have limitations related to reliability and task range. We previously introduced a visual-search (VS) model for predicting human-observer performance at detection-localization tasks in PET and SPECT. With this model, a holistic search locates suspicious image blobs, which are then analyzed by a statistical discriminant. We tested several approaches to identifying blob features for the holistic search, including a parametric blob model based on EM implementation of a Gaussian mixture model (GMM). Comparison of several VS models against a scanning model observer and human observers was carried out with a localization ROC (LROC) study based on simulated SPECT lung images. RBI reconstructions performed with either attenuation correction or resolution compensation were postsmoothed with a 3D Gaussian filter. The study objective was to identify the optimal level of postsmoothing given a fixed number of iterations. A channelized nonprewhitening discriminant for both the scanning observer and the analysis phase of the VS models. The VS models were considerably more accurate than the scanning model at predicting human performances. A high number of blobs identified in the holistic search could not be fit by the EM-GMM algorithm. For the successful fits, the number of blob pixels corresponded well with blob area as calculated from the Gaussian covariance matrix. Overall, the reliability of the VS model will depend on how the holistic search is carried out. The high failure rate with the EM-GMM approach might be remedied with better initialization.