We evaluated undersampling in MRI using a multicoil SENSE reconstruction with no regularization based the detection of signals by humans. We used a sparse difference-of-Gaussians (S-DOG) model to predict human performance in the detection of a small and large signal in anatomical backgrounds. The prediction was then validated using human observer two-alternative forced choice (2-AFC) tasks. Our model predicted a decrease in performance for both the small and large signal from 4X to 5X acceleration. Our observer study validated that prediction. This approach may lead to a way of assessing image quality that predicts human performance with fewer observer studies.