The Problem of Hidden Noise in MR Image Reconstruction

The performance of modern image reconstruction methods is commonly judged using quantitative error metrics like mean squared-error and the structural similarity index, where these error metrics are calculated by comparing a reconstruction against fully-sampled reference data. In practice, this reference data contains noise and is not a true gold standard. In this work, we demonstrate that this “hidden noise” can confound performance assessment methods, leading to image quality degradations when typical error metrics are used to tune image reconstruction performance. We also demonstrate that a new error metric, based on the non-central chi distribution, helps resolve this issue.