Data-Driven Multi-contrast Spectral Microstructure Imaging with InSpect

Abstract

We introduce and demonstrate an unsupervised machine learning method for spectroscopic analysis of quantitative MRI (qMRI) experiments. qMRI data can support estimation of multidimensional correlation (or single-dimensional) spectra, which allow model-free investigation of tissue properties, but this requires an ill-posed calculation. Moreover, in the vast majority of applications ground truth knowledge is unobtainable, preventing the application of supervised machine learning. Here we present a new method that addresses these limitations in a data-driven way. The algorithm simultaneously estimates a canonical basis of spectral components and voxelwise maps of their weightings, thereby pooling information across whole images to regularise the ill-posed problem. We show that our algorithm substantially outperforms current voxelwise spectral approaches. We demonstrate the method on combined diffusionrelaxometry placental MRI scans, revealing anatomically-relevant substructures, and identifying dysfunctional placentas. Our algorithm vastly reduces the data required to reliably estimate multidimensional correlation (or single-dimensional) spectra, opening up the possibility of spectroscopic imaging in a wide range of new applications.

Combined diffusion‐relaxometry microstructure imaging: Current status and future prospects

“… Data‐driven regularization, where a fixed number of correlation spectra are assumed within the image. 117 , 118 This approach is related to the previously mentioned blind source‐separation methods 105 , 106 , 107 and seeks a lower‐dimensional spectral representation of the image that is supported by the data, effectively regularizing the inversion by sharing information across voxels. This approach is appropriate when seeking to discover prominent microstructural features, at the expense of estimating spectra in every voxel.…”

“…The proportion of each voxelwise spectra that lies within each of these prominent regions is then calculated (hence the name spectral integration), yielding scalar indices often termed apparent spectral volume fractions. The data‐driven regularization methods described above 117 , 118 provide an alternative approach for deriving maps.…”

“…[ 152 , 153 ]) or unsupervised learning (eg, Ref. [ 118 ]), as ground truth information on tissue microstructure is not typically obtainable.…”

“…This approach has been used in combined diffusion‐relaxometry, 100 but was also used in earlier 1D multicomponent exponential modeling applications 115,116 . Estimation theoretic analysis can be used to show that spatial smoothness constraints can theoretically reduce the ill‐posedness of the ILT by orders of magnitude in both 1D 116 and higher‐dimensional 55 settings. Data‐driven regularization, where a fixed number of correlation spectra are assumed within the image 117,118 . This approach is related to the previously mentioned blind source‐separation methods 105‐107 and seeks a lower‐dimensional spectral representation of the image that is supported by the data, effectively regularizing the inversion by sharing information across voxels.…”

“…[152,153]) or unsupervised learning (eg, Ref. [118]), as ground truth information on tissue microstructure is not typically obtainable. There are a wide range of diffusion MRI brain microstructure imaging techniques, many of which show promise in the clinic.…”