PyQMRI: An accelerated Python based Quantitative MRI toolbox

Abstract

Various medical examinations are seeing a shift to a more patient centric and personalized view, based on quantitative instead of qualitative observations and comparisons. This trend has also affected medical imaging, and particularly quantitative MRI (qMRI) gained importance in recent years. qMRI aims to identify the underlying biophysical and tissue parameters that determine contrast in an MR imaging experiment. In addition to contrast information, qMRI permits insights into diseases by providing biophysical, microstructural, and functional information in absolute quantitative values. For quantification, biophysical models are used, which describe the relationship between image intensity and physical properties of the tissue for certain scanning sequences and sequence parameters. By performing several measurements with different sequence parameters (e.g. flip angle, repetition time, echo time) the related inverse problem of identifying the tissue parameters sought can be solved. Quantitative MR typically suffers from increased measurement time due to repeated imaging experiments. Therefore, methods to reduce scanning time by means of optimal scanning protocols and subsampled data acquisition have been extensively studied. However, these approaches are typically associated with a reduced SNR, and can suffer from subsampling artifacts. To address both aspects, it has been shown that the inclusion of a biophysical model in the reconstruction process leads to much faster data acquisition, while simultaneously improving image quality. The inverse problem associated with this special reconstruction approach requires dedicated numerical solution strategies (Block et al. known as model-based reconstruction. Model-based reconstruction is based on variational modeling, and combines parallel imaging and compressed sensing to achieve acceleration factors as high as 10x the speed of fully sampled acquisitions. The method directly solves for the unknown parameter maps from raw k-space data. The repeated transition from k-space to image-space, combined with the involved non-linear iterative reconstruction techniques to identify the unknown parameters, often leads to prolonged reconstruction times. This effect gets even more demanding if 3D image volumes are of interest.