Learning a nonlinear channelized observer for image quality assessment

We propose two algorithms for task-based image quality assessment based on machine learning. The channelized Hotelling observer (CHO) is a well-known numerical observer, which is used as a surrogate for human observers in assessments of lesion detectability. We explore the possibility of replacing the linear CHO with nonlinear algorithms that learn the relationship between measured image features and lesion detectability obtained from human observer studies. Our results suggest that both support vector machines and neural networks can offer improved performance over the CHO in predicting the human-observer performance.