Feature selection for learning-machine numerical observer

It is now accepted that image quality should be evaluated using task-based criteria, such as human-observer (HO) performance in a lesion-detection task. Because an HO study is costly and time consuming, the development of a numerical observer (NO) surrogate is highly desirable. NO, like the channelized Hotelling observer (CHO), typically uses some features, i.e. numerical values, extracted from images to predict HO performance. Recently, we proposed and successfully tested a supervised-learning approach for modeling HOs with a machine-learning algorithm (namely a support vector machine). In the supervised-learning approach the goal is to identify the relationship between measured image features and HO defect likelihood scores. In this work we further explore the proposed learning approach by evaluating the image feature selection. Our preliminary results use, as a starting point, the image features as those used in CHO methodology, namely the outputs of four constant-Q frequency-band filters intended to model the human visual system, indicating that the features have significant influence on the NO accuracy in predicting HO performance.