Some commands in MRtrix3 have substantial RAM requirements, and can therefore fail even on a relatively modern machine; e.g.:
tcksift2must store, for every streamline, a list of every fixel traversed, with an associated streamline length through each voxel.
fixelcfestatsmust store a sparse matrix of fixel-fixel connectivity between fixels in the template image.
In both of these cases, the memory requirements increase in proportion to the number of streamlines (directly proportionally in the case of SIFT/SIFT2, less so in the case of Connectivity-based Fixel Enhancement (CFE)). They also depend on the spatial resolution: If the voxel size is halved, the number of unique fixels traversed by each individual streamline will go up by a factor of around 3, with a corresponding increase in RAM usage in SIFT/SIFT2; but the total number of unique fixels increases by up to 8, and hence the total number of possible fixel-fixel connections goes up by a factor of 64! The RAM usage of CFE therefore increases by a substantial amount as the resolution of the template is increased. Unfortunately in both of these cases it is likely theoretically impossible to reduce the RAM requirements in software any further than has already been done; the information stored is fundamental to the operation of these algorithms.
In both cases, the memory usage can be reduced somewhat by reducing the number of streamlines; this can however be detrimental to the quality of the analysis. Possibly a better solution is to reduce the spatial resolution of the underlying image, reducing the RAM usage without having too much influence on the outcomes of such algorithms.
For SIFT/SIFT2, the subject FOD image can be down-sampled using e.g.:
$ mrresize in.mif out.mif -scale 0.5
Note that it is not necessary to use this down-sampled image for tractography, nor for any other processing; it is simply used for SIFT/SIFT2 to reduce memory usage. Additionally, by performing this down-sampling using MRtrix3 rather than some other software, it will ensure that the down-sampled image is still properly aligned with the full-resolution image in scanner space, regardless of the image header transformation.
For CFE, it is the resolution of the population template image that affects the memory usage; however using higher-resolution images for registration when generating that population template may still be beneficial. Therefore we advocate downsampling the population template image after its generation, and otherwise proceed with Fixel-Based Analysis (FBA) using this down-sampled