Global intensity normalisation

Most DWI models derive quantitative measures by using the ratio of the DW signal to the b=0 signal within each voxel. This voxel-wise b=0 normalisation implicitly removes intensity variations due to T2-weighting and RF inhomogeneity. However, unless all compartments within white matter (e.g. intra- and extra-axonal space, myelin, cerebral spinal fluid (CSF) and grey matter partial volumes) are modelled accurately (i.e. with appropriate assumptions/modelling of both the compartment diffusion and T2), the proportion of one compartment in a voxel may influence another. For example, if CSF partial volume at the border of white matter and the ventricles is not taken into account, then a voxel-wise normalisation performed by dividing by the b=0 (which has a long T2 and appears brighter in CSF than white matter in the T2-weighted b=0 image), will artificially reduce the DW signal from the white matter intra-axonal (restricted) compartment, ultimately changing the derived quantitative measures.

In our previous work investigating differences in Apparent Fibre Density (AFD) we opt to perform a global intensity normalisation between subjects. This avoids the aforementioned issues, but also comes with its own set of challenges and assumptions inherent to specific strategies to deal with this. Aside from the problem of how to select the region to perform global intensity normalisation (that is unbiased with respect to the groups in the analysis), the data must also be bias field corrected first, to eliminate low frequency intensity inhomogeneities across the image.

In theory, a sound approach to global intensity normalisation would be to normalise based on the median CSF b=0 intensity across all subjects (on the assumption that the CSF T2 is unlikely to be affected by pathology). However, in practice it is surprisingly difficult to obtain a robust partial-volume-free estimate of the CSF intensity due to the typical low resolution of DW images. For participants less than 50 years old, due to reasonably small ventricles, it can be difficult to identify pure CSF voxels at 2-2.5mm resolutions. We therefore recommend performing a global intensity normalisation using the median white matter b=0 intensity. While the white matter b=0 intensity may be influenced by pathology-induced changes in T2, our assumption is that such changes will be local to the pathology and therefore have little influence on the median b=0 value.

We have included the dwiintensitynorm script in MRtrix to perform an automatic global normalisation using the median white matter b=0 value. The script input requires two folders: a folder containing all DW images in the study (in .mif format) and a folder containing the corresponding whole brain mask images (with the same filename prefix). The script runs by first computing diffusion tensor Fractional Anisotropy (FA) maps, registering these to a study-specific template, then thresholding the template FA map to obtain an approximate white matter mask. The mask is then transformed back into the space of each subject image and used in the dwinormalise command to normalise the input DW images to have the same b=0 white matter median value. All intensity normalised data will be output in a single folder. As previously mentioned all DWI data must be bias field corrected before using dwiintensitynorm, for example using dwibiascorrect.

As an alternative to the dwiintensitynorm script, we have recently provided a new command called mtnormalise, which performs multi-tissue informed intensity normalisation in the log-domain. The benefit of the mtnormalise command is that normalisation can be performed independently on each subject, and therefore does not require a computationally expensive registration step to a group template. However, to perform multi-tissue CSD with 3 tissue types (WM, GM, CSF) currently requires DWI data with multiple b-values (this will change at some stage, when the implementation of single-shell 3-tissue CSD becomes available).