SVM-based visual-search model observers for PET tumor detection

Many search-capable model observers follow task paradigms that specify clinically unrealistic prior knowledge about the anatomical backgrounds in study images. Visual-search (VS) observers, which implement distinct, feature-based candidate search and analysis stages, may provide a means of avoiding such paradigms. However, VS observers that conduct single-feature analysis have not been reliable in the absence of any background information. We investigated whether a VS observer based on multifeature analysis can overcome this background dependence. The testbed was a localization ROC (LROC) study with simulated whole-body PET images. Four target-dependent morphological features were defined in terms of 2D cross-correlations involving a known tumor profile and the test image. The feature values at the candidate locations in a set of training images were fed to a support-vector machine (SVM) to compute a linear discriminant that classified locations as tumor-present or tumor-absent. The LROC performance of this SVM-based VS observer was compared against the performances of human observers and a pair of existing model observers.