What are the granularity levels and how to use them?

It is recommended to read Djamanikova et al, Neuroimage. 2014 Nov 1;101:168-76 about the multi-granularity analysis.

There are several reasons why we believe the multi-granularity analysis would be important;

Reason #1:
The images below show the brainstem and the cerebellum at five different granularity levels. With the highest granularity (Level 5), many structures in the brainstem are defined. One may think that the more structures are defined, the better. If we keep increasing the granularity, the highest level we can achieve is the single voxel and there would be more than 1 million voxels (structures) defined in the image. This is the voxel-based analysis (VBA), which is the most widely used quantitative analysis tool.

Compared to VBA, even Level 5 with currently 286 defined structures is a low-granularity analysis. However, we can argue even the granularity of Level 5 is too high for some brain areas such as the brainstem because T1-weighted images do not have enough contrasts to resolve all the defined structures in Level 5; the volume reports of the fine intra-brainstem structures are not reliable because they are invisible. These reports could be reliable if we use DTI because they are resolved in the DTI images, but not in T1-weighted images. If we are interested in the brainstem anatomy using T1, probably the granularity level at Level 3 or 4 would be appropriate.

The appropriate granularity level varies depending on the image contrasts and brain regions. It is straightforward to perform statistical analysis at all levels, but if statistical significance is found in a structure in one of the granularity levels, it is very important to confirm that such a structure is visible in the images and thus the result is reliable.

Head


Reason #2:
Suppose we find 10% atrophy in the hippocampus, which is statistically significant. This finding would be misleading if the entire temporal lobe has atrophy with the same amount. To test the spatial specificity, we need to confirm the absence of similar findings in the lower granularity levels.

Head


We provide two types of hierarchical trees, Type I and II, for the multi-granularity analysis. The structures in all levels are comprised of 286 structural units (Version 5L) and the hierarchical trees are created by combining them differently. Type I is based on ontology widely accepted by anatomists. For example, Level 1 / Type I divides the brain into telencephalon, diencephalon, mesencephalon, metencephalon, and myelencephalon. As shown in the above figure, the metencephalon comprised of the pons and the cerebellum combined. Type II hierarchy is based on the convention widely used in radiology. For example, Level 1 divides the brain into the hemispheres, the brainstem, and the cerebellum. Another difference is how the gray matter and white matter are separated. In Type I, they are separated early on (at Level 2) while they are not separated in Type II. As shown above, at Level 3, the five lobes are divided both in Type I and II, but Type II reports lobar volumes with the gray and white matter combined.

Similar to genetic analysis, in which information from the 3 billion base pair needs to be reduced to, for example, several thousand protein-encoding genes, granularity reduction from more than 1 million voxels to a much smaller number of structural representations could provide different views about anatomical features. The problem is, even if we reduce the voxel information into, for example, 200 structures, from a purely mathematical point of view, there are infinite ways to combine voxels to define the 200 structures. The conventional notion of "brain structures" is one of the most natural ways to define structures, but still there are numerous criteria for structural definitions. We try to provide a comprehensive set of pre-defined structures, but users can devise their own way to define structures based on their hypothesis by modifying the hierarchical relationship defined in the "Multilevel_lookup_table.txt."

Please note that the same structure names in different levels may not have the same definition.

For example, the "hippocampus" in Type 1 - Level 4 is a combination of the "hippocampus" and "fimbria" in Type 1 - Level 5. Consequently, the volumes of the hippocampus in Level 4 and Level 5 are not the same.

Structures should always be referred with the Type and Level, like "Hippocampus (Type1/Level4)."