Adaptive Feature Selection for Model Observers: Reducing Reliance on Prior Knowledge for Task-based Assessment

Mathematical model observers that are applicable for clinically realistic tasks are of particular interest for task-based assessments. We propose an efficient search-capable model observer that can operate without explicit background knowledge. In place of existing scanning-observer frameworks that invoke background subtraction, the model generates adaptive binary discriminants from feature data containing implicit background information. Initial validation of the model against human-observer data from a PET localization ROC (LROC) study is presented.