Myelin mapping is of great interest for neuroimaging, particularly in studies of brain development. However, methods to image myelin are typically time consuming and/or relatively noisy [1], making them impractical for use in applications such as neonatal imaging, where scan times are very limited. Here we propose a simple method for extracting maps of approximate myelin content from routinely acquired T1-weighted (T1w), T2-weighted (T2w) and proton density weighted (PDw) scans, and show promising results in neonatal scans.

The proposed approach is illustrated in Figure 1. We model the tissue as three compartments: tissue water, free water, and lipid; in the brain, the lipid compartment should correspond to myelin. While there will undoubtedly be some exchange between these compartments, we assume that the impact of this exchange is negligible, so that the signals from each compartment can be assumed independent of each other. We further assume that the relaxation rates of these compartments are invariant across the brain. Hence, for a particular combination of T1w, T2w and PDw sequences, the MRI signal from each compartment will scale linearly with its volume fraction, giving rise to the matrix equation:

$$ s = H f $$

where $$$s$$$ is the vector of T1w,T2w & PDw signals, $$$H$$$ is the matrix of expected T1w, T2w & PDw signals for each compartment, and $$$f$$$ is the vector of volume fractions per compartment. For each subject, the matrix $$$H$$$ was obtained by empirically measuring signals from suitable locations in the data themselves. For lipids, measurements were taken from subcutaneous fat. For tissue water, measurements were taken from the (non-central sulcus) cortex (since it is cell-dense and assumed myelin-free in this age range). For free water, measurements were taken from the ventricles. Volume fractions $$$f$$$ were then estimated using a non-negative least-squares (NNLS) algorithm (this helps to constraint the problem by ensuring positive volume fractions).

Data were collected on a 3T Philips Achieva (Best, Netherlands) from lesion-free infants born pre-term scanned at term-equivalent age, and consisted of a T1-weighted MP-RAGE sequence (TE/TR/TI=4.6/17/1500ms, flip angle=13˚) and dual-echo T2-weighted FSE sequence (TE1/TE2/TR=12.8/160/8700ms), with parental consent and approval from the local Ethics Committee. The two echoes from the T2-weighted sequence were used as the PDw and T2w image respectively. For each subject, the T1w images were coregistered to the T2w images and processed as described above.

Figures 2-4 show results from three subjects, showing for each subject the volume fraction of each compartment. The free and tissue water compartments are well characterised, with the lipid signal having much lower intensity as expected in this cohort. Encouragingly, the lipid compartment displays a spatial pattern consistent with the expected pattern of myelination in this age range [2], and is consistent between the different subjects, supporting the interpretation of the lipid signal within the brain as a marker of myelin. Note that the white matter appears as a mixture of tissue and free water, which agrees well with the high mean diffusivity values reported in this age range [2]. There is also evidence of residual artefacts due to misregistration manifesting as a band of apparent lipid signal at the edges of the cortex, particularly in the superior aspects of the brain. Nonetheless, there is evidence of increased lipid volume fraction within the central sulcus in all 3 subjects, over and above these misregistration-induced artefacts, consistent with the early myelination of the corticospinal tracts in this age range [2].

The simple method proposed here provides a fast and stable means of deriving semi-quantitative maps of myelin from routinely acquired data, making it suitable for use in a wide variety of contexts. Of note, this may be of use in situations where suitable data are collected as part of a routine clinical protocol, but dedicated myelin imaging methods are impractical due to scan time constraints (this is the case particularly for neonatal imaging).

The approach relies on two assumptions: that relaxation rates are invariant per compartment; and that the effects of exchange are negligible. These assumptions are most likely simplistic, implying that the results are undoubtedly biased to some extent. Future work will focus on estimating the impact of violations of these assumptions on the results, as well as improving image registration and per-compartment signal calibration. Despite these concerns, the method provides a map that appears both sensitive and specific to myelin, and shows promise for applications such as studies of brain development.