Re: Autoseg test

[Rick Giuly]

Are you binning down also? I think that has a good chance of helping.
It'll look at an 11x11 window of pixels. So when you bin down, it can
effectively see more inside of that window. Also, considering the
resolution, it appears most of the texture is still there after binning.

Nice find with the command, I'll add that to the docs.

-Rick

Alex Perez wrote:
> Hey Rick,
>
> Ok, cool. I downloaded the new update. I'm currently running two
> processes, one with countourListThreshold=0.6 and the other =0.8.
>
> I adjusted the contrast from last time, maybe it will help.
>
> Also, I figured out that you can create the thresholded images using:
>
> imodmop -mask 1 trainingData.mod trainingData.mrc thresh.mrc
>
> This will do it automatically for you, without needing to use fiji or
> anything else.
>
> Thanks,
> Alex
>

[Rick Giuly]

Ok

You'll want both steps usually

If you just do step one you get the voxel results but not the "blobs" folder

Rick

On Nov 1, 2012 5:59 PM, "Alex Perez" <a3p...@ucsd.edu> wrote:

Yea, I tried it binned by two.

One question...should I still be running it with both the --step1 and --step2 arguments? Or just --step2?

[Rick Giuly]

Hi Alex, that's cool. Glad to hear it.

Another thing that helps is balancing the examples. When the program
starts running you will see that it starts listing out how many mito
examples and how many "other" examples it's finding. Those should be
about the same number. The way to adjust them is with the weight parameters:
for example:
voxelWeights=0.0130,0.00064

Right now mito count is higher I think, so the first number should be
brought down proportaionally.. something like voxelWeights=0.0090,0.00064

Just another thing to try. You'll see the results of modifying this
weight most clearly in the voxel output (as opposed to the blob output).

-Rick


Alex Perez wrote:
> Ok, it seems I'm having the best results with --contourListThreshold <
> 0.2. I just did a bunch of runs with it set to 0.2, 0.1, 0.05.
>
> There aren't any false positives with 0.2, but it does miss a few mitos.
> When I take it below 0.2, there start to be false positives,
> particularly with myelin.
>
> This was with data binned by 2 and with the contrast enhanced. It seems
> the contrast enhancement did help a good amount. I'll keep playing around...
>
>

[Rick Giuly]

The last one in the list is the one you want to use, it's the "total"

-Rick


On 11/02/2012 03:32 PM, Alex Perez wrote:

For the mito vs other count, I see that it lists it out a bunch of times (different iterations?). For example:

154 out of 175
recorded example counts so far: {'mitochondria': 487, 'other': 333}
155 out of 175
recorded example counts so far: {'mitochondria': 492, 'other': 334}
156 out of 175
recorded example counts so far: {'mitochondria': 498, 'other': 335}
recorded example counts: {'mitochondria': 504, 'other': 335}

Is it those last two numbers I want to use? So in this case, mito=504, other=335?

On Fri, Nov 2, 2012 at 2:52 PM, Alex Perez <a3p...@ucsd.edu> wrote:

Ok cool, I'll give it a shot. Do you think having larger training data would help as well?