Option to prune small twigs before nblast

[jefferis]

By combining strahler order and segment length we could probably remove a large number of small terminal branches without any impact on nblast results.

[jefferis]

maybe reverse Strahler order

[jefferis]

Could also remove anything tagged as not a branch

[jefferis]

Here is a first attempt at this. Trials on some DA1 PNs show a reduction in the number of nodes of about 25%, which seems useful though not huge. It may have a greater impact on neurons with very extensive arbours.

prune_terminals <- function(x, maxorder=1, maxlength=2000,...) {
  badsegs = strahler_order(x)$segments <= maxorder &
    seglengths(x) <= maxlength
  goodids = unlist(x$SegList[!badsegs])
  prune_vertices(x, goodids, invert = T)
}

> da1s.p=nlapply(da1s, prune_terminals)
> da1s.p.dps=dotprops(da1s.p/1e3, k=5, resample=1)
> da1s.dps=dotprops(da1s/1e3, k=5, resample=1)

> summary(da1s.p.dps)/summary(da1s.dps)
            nodes cable.length
27295   0.6759853    0.6759853
57311   0.7644979    0.7644979
57323   0.7818021    0.7818021
57353   0.8527709    0.8527709
57381   0.6872659    0.6872659
61221   0.7637326    0.7637326
755022  0.7949153    0.7949153
2863104 0.7189805    0.7189805

> summary(summary(da1s.p.dps)/summary(da1s.dps))
     nodes         cable.length   
 Min.   :0.6760   Min.   :0.6760  
 1st Qu.:0.7111   1st Qu.:0.7111  
 Median :0.7641   Median :0.7641  
 Mean   :0.7550   Mean   :0.7550  
 3rd Qu.:0.7851   3rd Qu.:0.7851  
 Max.   :0.8528   Max.   :0.8528  
> 

[jefferis]

mvp2=read.neuron.catmaid(2333007)
mvp2.orig=dotprops(mvp2/1e3, k=5, resample=1)
mvp2.p=dotprops(prune_terminals(mvp2)/1e3, k=5, resample=1)
> nvertices(mvp2.orig)
[1] 29513
> nvertices(mvp2.p)
[1] 10959

So more significant, but it actually takes 30 s to do the whole dotprops and strahler order thing. More than 20s of that is dotprops, with much of the time spent rbinding.

[jefferis]

NB have been having a problem where neurons have cycles

> prune_terminals(da1s[[1]])
Error in as.seglist.igraph(masterg <- x, origin = origin, Verbose = Verbose) : 
  Graph has cycles!
Called from: as.seglist.igraph(masterg <- x, origin = origin, Verbose = Verbose)

[jefferis]

Looks like this relates to a bug introduced into nat when trying to speed up resampling of neurons in jefferis/nat@77ca3e51730af1d5102dc5baf625c352f0504462.

[jefferis]

We now have a good implementation of this in prune_twigs: https://github.com/jefferis/elmr/blob/e34dc5e4dadc93160dba7889031501462bf9b966/R/prune_twigs.R#L1-L20

(good enough that we may move to nat).

[jefferis]

So we can update nblast_fafb to use this by default. @schlegelp would this have any impact on CATBOT?

[schlegelp]

No. Catbot uses the basic nblast function and does all the preprocessing itself. It currently uses the pre-processed simp10.dps overnight dump. If you are changing nblast_fafb, should I change that behaviour to using the twigs5.dps to match nblast_fafb?

[jefferis]

I think twigs5 is def better than simp10 although it might take longer.