I am trying to track the auditory radiation (medial geniculate nucleus to Heschl’s gyrus) with 5 shell data (max b-val 3000) but have been running into a lot of difficulty because it doesn’t seem like there are peaks in the direction of the auditory radiation (where I know it should be anatomically). I preprocessed with Denoising (MP-PCA), Gibbs ringing correction, Eddy, motion, and distortion correction, Bias field correction, and Brain extraction (FSL BET). I used Dhollander for response function estimation, CSD for FOD calculation, and probabilistic tracking (default).
Are there any ideas as to what might be causing the lack of peaks in the direction of the auditory radiation? Is this a data resolution issue, or is it possible I’m doing something incorrectly in the preprocessing or tracking? I am able to track the full brain well and extract structures—it seems to be the auditory radiation that’s giving me trouble.
It’s a very general and challenging question to tackle. I don’t have any experience with this particular tract so I don’t know how dense it is and therefore whether it is simply too small in volume to be visible in the DWI signal. Bear in mind that a bundle of axons has to occupy a reasonable fraction of the volume of an image voxel (say ~5%) for it to be possible to distinguish the DWI signal emanating from that bundle from thermal noise. There are many more inter-areal connections in the brain than network neuroscientists would have you believe, but many of them consist of a very small number of axons, and so won’t even appear in FODs let alone tractography.
If it is potentially an issue of thermal noise deep in the brain you could have a go at my new denoising implementation. No guarantee that it will magically make your tract appear out of nowhere, but if you’re failing at the very first hurdle with the known tract trajectory not being visible in the FODs, might as well give yourself the best chance possible.
I would also suggest downloading some HCP data and seeing if you can identify FOD lobes corresponding to that bundle trajectory. Will give a better sense of what an improvement in data acquisition quality might contribute.
Thank you for the reply! I will definitely give dwidenoise2 a try and see if I am able to reconstruct the peaks. I have tried with HCP data, and I was able to reconstruct it there (more plausibly than in my own data).
Our acquisition is fairly good resolution (3T scanner, 96 directions, 1.7 mm isotropic resolution), but I am looking into upsampling the data. I think a registration-based approach is better for my data—would this be done with mrgrid using a T1 as the template?
I consider “upsampling” and “registration” to be two very different operations, so it’s not clear to me if you’re making reference to something specific or making an erroneous conflation.
Upsampling the DWI data may only help in the specific instance where you have a bundle that undergoes a lot of curvature. When tractography performs interpolation of the SH coefficients, it might fail to properly follow the curve from the bundle’s orientation in one voxel to the orientation of that same bundle in the adjacent voxel. Here, upsampling the DWI data, and therefore performing CSD on interpolated DWI data, may yield an intermediate FOD that appropriately reflects the change in fibre orientation of that bundle, and therefore improve probability of success of tractography. But upsampling DWI data will not magically make FOD peaks appear where previously they were not present.
If you go down the road of upsampling, using the T1-weighted image voxel grid is one option, but it’s not the exclusive answer. You can use any template voxel grid, and you can specify any spatial resolution you want. But there will be diminishing returns beyond a factor of 2.
I see—that makes sense. The registration-based upsampling I was referring to is a method I found that uses the T1 as a template (I believe it’s the T1-weighted voxel grid you’re referring to). Can this be done with mrgrid?
I don’t know how to interpret the phrase “registration-based upsampling”.
You can indeed use mrgrid to resample DWI data on the T1-weighted image voxel grid. But I would not call this “registration-based upsampling”: it’s just upsampling onto the voxel grid of another image, to which you may have previously registered. I don’t see the internal operation of that upsampling being “based on” the outcomes of registration. It might be that this is embellishing language used by a third party. But you could define any arbitrary voxel grid of the same spatial resolution as the T1-weighted image, and resample your DWI data onto that grid, and the benefits of such upsampling would be equivalent (with the exception of any downstream process that explicitly requires that DWI and T1w image data be defined on exactly the same voxel grid).