How Should We Evaluate Uncertainty in Accelerated MRI Reconstruction?

Reconstructing accelerated MRI is an ill-posed problem. Machine learning has recently shown great promise at this task, but current approaches to quantifying uncertainty focus on measuring the variability in pixelwise intensity variation. Although these provide interpretable maps, they lack structural understanding and they do not have a clear relationship to how the data will be analysed subsequently. In this paper, we propose a new approach to evaluating reconstruction variability based on apparent anatomical changes in the reconstruction, which is more tightly related to common downstream tasks. We use image registration and segmentation to evaluate several common MRI reconstruction approaches, where uncertainty is measured via ensembling, for accelerated imaging. We demonstrate the intrinsic variability in reconstructed images and show that models with high scores on often used quality metrics such as SSIM and PSNR, can nonetheless display high levels of variance and bias in anatomical measures.