Learning to Predict Error for MRI Reconstruction

In healthcare applications, predictive uncertainty has been used to assess accuracy. this paper, we demonstrate that estimated by the current methods does not highly correlate with prediction error decomposing latter into random and systematic errors, showing former is equivalent variance of error. addition, observe unnecessarily compromise performance modifying model training loss estimate target jointly. We show estimating them separately without modifications improves performance. Following this, propose a novel method estimates labels magnitude in two steps. on large-scale MRI reconstruction task, achieve significantly better results than state-of-the-art estimation methods.