Hello,
I had a question of a special use case.
I want to simulate a lesion on a healthy DWI image.
I want to give a healthy DWI image a lesion so that I can compute tractography on it.
Is there a way to intersect the DWI image with a lesion mask in DWI space using MRTrix? What commands could I use?
I tried searching on the online forums but couldn’t quite find what I was looking for.
Your help is appreciated.
Thank you
Best,
Apoorva
Hi @apoorvakelkar,
With a DWI image and a lesion image that is 0s where lesion is and 1 everywhere else, use mrmath (mrmath — MRtrix3 3.0 documentation) to multiply the images together.
mrmath dwi.nii.gz lesion_0_mask.nii.gz product dwi_0_lesion_mask.nii.gz
Best,
Steven
I do not think this approach will make a very realistic injury. Why not use actual stroke datasets. The ARC provides hundreds of diffusion weighted images suitable for DTI analyses from individuals with chronic stroke. The SOOP provides thousands of acute DWI scans. For details see here and here.
Hello @neurolabusc,
Thank you for the datasets. It is helpful to know of these publicly available datasets.
We’ve been evaluating these options further and haven’t encountered any specific method that satisifies what we would prefer to do.
Are you aware of any technique that simulates natural lesions on healthy brain images that also preserves the directional information needed to produce diffusion tractography?
Thank you
Best,
Apoorva
@apoorvakelkar your question is underspecified. Maybe you need to describe your objective, a concrete example of your preliminary attempts, the methods you are familiar with and your level of expertise. Consider the DWI data from the OpenNeuro ARC participant sub-M2307. Note that in addition to the diffusion images this repository includes the bvec and bval files to describe the gradient direction and magnitude. You could analyze this diffusion data with your favorite pipeline (AFNI, MRtrix, FSL) or nii_preprocess which adapts FSL for stroke lesions.
@neurolabusc Thanks for your thoughts so far Chris, I figured I would reply after @apoorvakelkar updated me! I don’t think we’ve met yet personally and we appreciate it.
Put simply, we are interested in examining the effects of lesions warped from real stroke cases to matched controls on resulting tractography. We have several specific questions about the effects of strokes on network topology and specific tracts in simulated cases. As soon as we started to think more about this problem, several options started to occur to us, some within our scope of expertise, and some beyond it.
We had started to consider what a “realistic” lesion could mean when working directly with the bval/bvec files within a stroke-to-control warped lesion mask. However, that seems a bit fraught because asking to move the DWI directions and magnitudes and obtain exact fiber estimates/FODs from person 1 to person 2 is a more complex transformation even before considering the real content and edges of the lesion masks (e.g., at any boundary of an ROI or edge of a lesion mask we might expect any number of abrupt directional/intensity changes across voxels).
So at the moment we have been evaluating a couple options. One that we are refining now is warping the lesions to controls, then observing the effects on tractography after exclusive masking, which of course has its own limitations. Maybe it’s also worth a shot just attempting the direct bval/bvec imputation as part of the lesion contents and observing the direct effect on the tractography.
In any case, these are helpful leads and we value your thoughts!
John Medaglia
You can certainly create synthetic lesions by inserting tissue from injured brains into scans from normal adults. A seminal example of this method was Brett et al. who used it to evaluate how the presence of a lesion impacted spatial normalization. Taylor Hanayik provides a nice Python script for making syntheitc lesions for T1 scans. However, I think the challenge with DWI data is working out the extent of the lesion, in particular if you are interested in the chronic case, where necrosis and Wallerian degeneration is present and so there is a lof of deformation of fibers as well as the packing of bundles, which can have distant effects on diffusion measures, as seen in McKinnon et al.. Unlike T1 images, diffusion images show EPI and eddy current distortions, that are dependent on many factors. Beyond this, the BIDS standard stores bvec’s in image space (not world space), so even an identical sequence will require rotating these values based on slice angulation and the distortions will differ. I certainly think you can model the effect of these effects, but a core concern becomes whether your findings reflect the assumptions of the model or the analyses methods.
If you are interested in acute impacts, the good news is that DWI is part of the standard of care, so you are likely able to get access to many pre-exisiting real world datasets. However, admission scans are focused on speed not quality, and derived values like TRACE and ADC are of interest. Therefore, these images often have intense EPI/eddy artifacts and do not contain the individual directions. A nice example of this is our OpenNeuro SOOP dataset. Alternatively, one can look at research grade chronic scans like our OpenNeuro ARC, but these datasets are smaller and have chronic injuries will impact the extent of diffusion changes.
Contact me if you want to have a zoom meeting.
@neurolabusc This is great, I follow what you are saying - and I would definitely appreciate a zoom meeting to discuss the deeper issues here even while we are tinkering with our cruder variant. I’ll take it to e-mail!
MANAGED BY INCF