Voxel vs Node Analysis

[Aly Kotb]

Hello All,

Could any one please explain to me the difference between voxel-based and node-based analysis. Why some studies apply their deep/machine learning models to fMRI image of voxels, and other studies they establish a connectivity matrix based on ROIs and deal with it as a node-based problem?

In other words, what each kind of these analyses of is targeting?

Thanks, 

Ali.

[Glasser, Matthew]

There are three relevant spatial approaches to brain imaging analysis:

1. Voxelwise: This is the oldest, but the most flawed.  Poor alignment methods across humans often fail to actually compare the same thing across individuals, basically making the whole thing a waste of time/source of false or limited findings in the literature.  A major goal of the HCP and related efforts has been to provide better alternatives to this approach.
2. Grayordinatewise: This is still an elementwise “dense” analysis, but uses much better methods of cross-individual alignment, including using surface vertices for the cerebral cortex (coming for hippocampus and cerebellum), which, ideally, are aligned using multi-modal information that is closely tied to cortical areas and functional networks. This will produce as good a result as is feasible, but will not address topological variability in cortical areas and functional networks.  A further extension of this sort of idea to non-topological alignment is hyperalignment.  For the deep white matter, the best analogy to this approach is using a white matter skeleton as in TBSS, and ideally using a multi-modal volume alignment that includes fiber orientation information (e.g., MMORF).  This volumetric alignment is also coming soon for the HCP Pipelines (and will also aid in analyses of globular subcortical structures).  There will still be some intervening superficial whitematter that will be hard to model using a dense approach because it is not represented on the cortical surface and there is topological variability in the cortical folds such that alignment in the volume is not possible. 
3. Imaging Derived Phenotypes or parcellated analysis: This approach focuses on the neuroanatomy of the brain, including cortical areas and sub areas, functional networks, and white matter tracts. Individualized delineations of these neuroanatomical structures enable representing topological variability across individuals.  Such analyses also have benefits in statistical sensitivity (averaging noise out across the neuroanatomical structures) and power (fewer statistical tests).  They are also more neurobiologically interpretable X structure is doing or affected by Y.

 

Work continues on the necessary methods, pipelines, and atlases to fully support the above conceptualization.

 

Matt.

--
You received this message because you are subscribed to the Google Groups "HCP-Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hcp-users+...@humanconnectome.org.
To view this discussion visit https://groups.google.com/a/humanconnectome.org/d/msgid/hcp-users/a7210642-aa59-44f6-8518-c61eff9189c6n%40humanconnectome.org.

 

---

The materials in this message are private and may contain Protected Healthcare Information or other information of a sensitive nature. If you are not the intended recipient, be advised that any unauthorized use, disclosure, copying or the taking of any action in reliance on the contents of this information is strictly prohibited. If you have received this email in error, please immediately notify the sender via telephone or return mail.

[Aly Kotb]

Thank you so much for this clarification. So, two more questions:

1- Why in functional parcellation papers use dense node representation like 10k or 32k per hemi-sphere not for example a 100x100 connectivity matrix based on one of the known Atlases?

2- Also, when to use thousands of nodes, and when to use a connectivity matrix based on a predefined ROIs atlas ?

Thanks,

Ali.

[Glasser, Matthew]

1. Maybe an example would help here.  If you are creating a new parcellation you would need to start from dense data.  Once you have a good parcellation, such as the HCP’s multi-modal parcellation, many downstream studies can use it.
2. There are times when folks essentially are just downsampling data, which you can do with lower resolution surface meshes or a “parcellation” of thousands of parcels.  In my view, it is better to use parcellations that are related to a particular level of the brain’s neurobiological organizational hierarchy, such as cortical areas.

The materials in this message are private and may contain Protected Healthcare Information or other information of a sensitive nature. If you are not the intended recipient, be advised that any unauthorized use, disclosure, copying or the taking of any action in reliance on the contents of this information is strictly prohibited. If you have received this email in error, please immediately notify the sender via telephone or return mail.