Re: NBLAST tangent calculation question

[Gregory Jefferis]

Dear Tom,

I basically always convert to vector cloud aka dotprops using only the vertex positions without segment connectivity. It’s simpler and the connectivity - at least when not smoothed - can be a bit jagged anyway. The key thing is to resample to a constant value order 1 um — this combined with k=5 seems to be a good compromise value given typical curvature of fly neurones. This does mean that it’s possible to include branches in that neighbourhood, but it doesn’t seem to have a big impact in practice. If you plot dotprops converted neutrons in R, you can actually appreciate that.

There is some code to compute scores using neurones directly, but as noted here:

http://jefferislab.github.io/nat.nblast/reference/WeightedNNBasedLinesetMatching.html

This was lightly tested.

Best wishes,

Greg. 

Sent from my iPhone

On 20 Jan 2018, at 22:31, Tom Kazimiers <kazim...@janelia.hhmi.org> wrote:

Hi Greg,

I am currently in the process of implementing some NBLAST functionality
into CATMAID. Part of this is computing tangents for each treenode of a
skeleton. I have a question regarding this and would appreciate it a lot
if you could point me in the right direction. The NBLAST paper says the
following about the used tangent computation:

"[…] the tangent vector (i.e. the local heading) of the neuron at each
point was computed as the first eigenvector of a singular value
decomposition (SVD) of the point and its 5 nearest neighbors."

In this particular paragraph light microscopy data is discussed and so I
wonder if you would calculate the tangent in the same way for already
skeletonized data like it exists in CATMAID or if it would also be okay
to just use a node's child and parent nodes, if any. I would imagine it
would yield better tangents if we only consider linked nodes.

Looking at the 5 nearest neighbors would potentially also include nodes
that are not connected to the target node. This might be something
unavoidable with LM data though (since no explicit skeletonization is
given). The nat R package source code also only revealed the 5 nearest
neighbor method.

Thanks,
Tom