Frechet distance for model observer training data selection

Abstract

In medical imaging, it has become widely accepted that image quality should be assessed using a task-based approach in which, for example, one evaluates human observer detection accuracy for a specific diagnostic task. These evaluations should be integral part of an imaging system optimization and testing. However, human observer studies with expert readers are costly and time-demanding. Consequently, model observers (MO) have been used as surrogates to predict human diagnostic performance. MOs use features derived from the images to accomplish these predictions. Some types of MOs require a set of data evaluated by humans for model tuning. In this work we present a methodology for tuning data selection. This selection is based on the Frechet distance between image-feature distributions. Specifically, in our experiments we show that MO, based on the Relevance Vector Machine (RVM), trained with the selected small subset of data has excellent performance in predicting human observer for diagnostic tasks.