The aim of this work is to improve the quality and contrast of synthetic FLAIR images by accounting for CSF partial volumes in quantitative relaxation time maps.Fluid Attenuated Inversion Recovery (FLAIR) imaging^1,2 is one of the most widely used sequences in neuroradiology. It provides T₂-weighted contrast with the suppression of cerebrospinal fluid (CSF) in the ventricles and sulci, allowing for examination of healthy and diseased tissues with short and moderate T₂ values. Relaxation time mapping methods quantify T₁, T₂, and proton density (M₀) in each voxel, opening the possibility to generate synthetic weighted images with arbitrary and retrospectively adjustable contrast. However due to partial volume contamination, effective relaxation times mapped in mixed-tissue voxels may be inaccurate, making the generation of weighted images calculated from the maps, including FLAIR images, problematic. In this work, we apply partial volume (PV) estimation³ with MR Fingerprinting (MRF)^4,5 to estimate and remove the partial volume signal contribution of CSF and generate improved synthetic FLAIR images from simultaneously acquired relaxation time maps.

MRF^4,5 was used to generate T₁, T₂, and M₀ maps (3000 images, resolution 1.17mmx1.17mm, imaging time 30s). A 7-component k-means analysis was performed on voxel T₁, T₂ pairs to identify regions of pure CSF, grey matter (GM), or white matter (WM) and partial volumes. Signal evolutions matching the T₁, T₂ centroids of the pure tissues were then entered into a 3-component PV sub-dictionary, $$$\mathbf{D}$$$.

PV estimation³ was applied to the MRF image series, in which each mixed voxel MRF signal evolution was modelled as a complex-valued weighted sum of component signal evolutions, assuming no exchange: $$s=\begin{bmatrix}w_{CSF} & w_{GM} & w_{WM} \end{bmatrix}\begin{bmatrix}d_{CSF} \\ d_{GM} \\ d_{WM} \end{bmatrix} = \bf w D$$

The partial volume weights were computed by multiplying the pseudoinverse of the PV sub-dictionary with each voxel signal evolution, $$$s$$$:

$$\mathbf {w}= s (\mathbf {D}^H \mathbf {D})^{-1} \mathbf {D}^H $$

The magnitudes of the complex weights were normalized to sum to one for visualization of tissue fractions.

The CSF sub-dictionary entry was multiplied voxel-wise by the complex CSF PV weight, and subtracted from the measured MRF signal to calculate CSF-removed MRF signal evolutions. MRF pattern matching was repeated to generate CSF-corrected relaxation time maps. The voxel synthetic FLAIR signal was calculated as $$s_{FLAIR} = |M_0| (1 - 2e^{-TI/T_1} + e^{TR/T_1}) e^{-TE/T_2}$$ with TI 2100ms, TR 15s, and TE 70ms.

Synthetic FLAIR images with and without CSF PV removed were compared with a FLAIR image from a clinically available protocol. The concept was demonstrated in one healthy volunteer and two patients scanned with MRF^4,5 at 3T (Skyra, Siemens Medical Solutions, Erlangen, Germany) under written informed consent in an IRB-approved study. The MRF data were reconstructed and processed offline in MATLAB (The Mathworks, Natick, Massachusetts, USA).

Figure 1 shows the estimated tissue fractions and demonstrates the effect of removing the fractional CSF signal on the mapped relaxation times. Changes occur in regions with CSF PV, with shorter voxel T₁ and T₂ in the remaining tissue, and reduced proton density in CSF regions.

Figure 2 shows synthetic FLAIR images for the healthy volunteer generated from MRF maps before and after CSF removal. In-plane flow from blood vessels is not removed in the MRF images, but contrast and visualization of the cortex and deep nuclei are improved.

Figure 3 shows synthetic FLAIR images in patients with multiple sclerosis and glioblastoma multiforme, with the closest available clinically-acquired FLAIR image provided for comparison. Contrast between normal-appearing GM and WM is improved, with reduced PV artifacts surrounding the ventricles, sulci, and lesions.

MRF maps the effective T₁ and T₂ in each voxel, causing partial volume with CSF and GM or WM to have values between those of the pure tissues. In T₂-weighted calculated FLAIR images, only pure fluid signals with long T₁ and T₂ are suppressed, leading to hyperintense signal in mixed voxels. Such hyperintensities reduce contrast between healthy and diseased tissues, and may be misinterpreted as pathology. These artifacts are not unique to MRF as a relaxation time mapping method. However, MRF has the unique possibility to resolve and remove partial volume quantities from the measured signal.

The assumption that the T₁ and T₂ of pure CSF, GM, and WM are known is justified by the presence of full voxels of each species, allowing for the relaxation times to be determined by k-means clustering.

MRF synthetic FLAIR image quality and contrast can be improved by estimating and removing the fractional signal originating from CSF. Future work includes improving flow compensation and the accuracy of the underlying relaxation time maps.