Retaining information from multidimensional correlation MRI using a spectral regions of interest generator

Abstract

Multidimensional correlation magnetic resonance imaging (MRI) is an emerging imaging modality that is capable of disentangling highly heterogeneous and opaque systems according to chemical and physical interactions of water within them. Using this approach, the conventional three dimensional MR scalar images are replaced with spatially resolved multidimensional spectra. The ensuing abundance in microstructural and chemical information is a blessing that incorporates a real challenge: how does one distill and refine it into images while retaining its significant components? In this paper we introduce a general framework that preserves the spectral information from spatially resolved multidimensional data. Equal weight is given to significant spectral components at the single voxel level, resulting in a summarized image spectrum. This spectrum is then used to define spectral regions of interest that are utilized to reconstruct images of sub-voxel components. Using numerical simulations we first show that, contrary to the conventional approach, the proposed framework preserves spectral resolution, and in turn, sensitivity and specificity of the reconstructed images. The retained spectral resolution allows, for the first time, to observe an array of distinct [Formula: see text]-[Formula: see text]-[Formula: see text] components images of the human brain. The robustly generated images of sub-voxel components overcome the limited spatial resolution of MRI, thus advancing multidimensional correlation MRI to fulfilling its full potential.

Figure 1

Overview of the proposed method, compared with the conventional approach. (A) Each voxel in a given image contains a 2D spectrum. Up to date, the only strategy to process the voxelwise spectra into images was to average them^,, manually identify spectral components, and generate sROIs accordingly (bottom panel). To circumvent the obvious limitations of the standard approach we propose here to first (B) identify all possible spectral peaks in each voxel, and then (C) apply a threshold and obtain binary voxelwise spectra, ${\mathbf{F}}_{\mathrm{bin}}$. (D) The binary spectra are then averaged to yield $⟨{\mathbf{F}}_{\mathrm{bin}}⟩$, (E) which is then used to generate the (F) sROIs. The binarization step ensures that even peaks with low prevalence in the image domain are represented in $⟨{\mathbf{F}}_{\mathrm{bin}}⟩$.

Figure 2

The simulations model. Each column shows a different simulated image-spectrum pair. Pixels in the inner most ring (I) contain all 5 spectral components (A–E), while pixels in the outer most ring (V) contain only a single spectral component (E). The synthetic phantom was created by superimposing all 5 image-spectrum pairs.

Figure 3

Numerical simulations results. The spectrum on the top row shows the 5 ground truth spectral components unweighted by their prevalence in the image, with their corresponding ground truth spatial images. Results from using the conventional approach to processing spectral imaging data are shown on the center row. $⟨\mathbf{F}⟩$ is simply an average spectrum across the entire image, and as such, it does not contain components A and B, and therefore cannot be used to correctly reconstruct the ground truth images. The bottom row shows the application of the suggested sROI generator, and its successful identification and reconstruction of the ground truth images.

Figure 4

Conventional processing and image reconstruction of multidimensional correlation MRI data. $T_1$−$⟨D⟩$, $T_2$−$⟨D⟩$, and $T_1$−$T_2$ distributions averaged across the image were used to locate the sROIs and to generate the signal fraction images (left to right). Note the loss of spectral resolution that resulted in redundancy of the $T_2$ dimension. Furthermore, many of the reconstructed images appear to be saturated, which implies the blending and smearing of the underlying spectral components.

Figure 5

Proposed framework for processing and image reconstruction of multidimensional correlation MRI data. $⟨{\mathbf{F}}_{\mathrm{bin}}^{T_1\mathrm{-}⟨D⟩}⟩$, $⟨{\mathbf{F}}_{\mathrm{bin}}^{T_2\mathrm{-}⟨D⟩}⟩$, and $⟨{\mathbf{F}}_{\mathrm{bin}}^{T_1\mathrm{-}{T}_2}⟩$ were computed using the proposed algorithm in Fig. 1 (top to bottom). These summarized spectra were then used to identify and define sROIs. Note that compared with the conventional approach, the sROI generator preserved more spectral components, has identified a significantly larger number of multidimensional peaks, and the spatial maps did not contain over-saturated region.