Tile Artifact

[Max Denena]

Hello,

I applied IsoNet to some tomograms of T cells, and I'm getting some kind of tiling effect that was not present before. Has anyone seen this before and would know how to fix it? I've attached two screenshots of the same slice through the tomogram for your reference.

Any help would be greatly appreciated.

-Max

[YUNTAO LIU]

Hi Max,

That is a well recognized problem of IsoNet. 

Two ways might be able to mitigate this problem.

1. increasing crop_size while decreasing cube_size in the predict step. This will inbtroduce more overlapping regions for each tile.

2. increasing the size of the mask before extraction, use larger values for "percentage", because artifacts often occur in the regions that are masked out from training.

I am working on a method to smooth out this artifact and still testing.

--
You received this message because you are subscribed to the Google Groups "IsoNet" group.
To unsubscribe from this group and stop receiving emails from it, send an email to isonet+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/isonet/a526da24-185d-470c-be40-6d9e79b6d8a1n%40googlegroups.com.

--

Best Regards,

Yuntao Liu,  Postdoc.

California NanoSystem Institute

University of California Los Angeles

[Max Denena]

That worked very well! Thank you for the help.

I tried a cube size of 32 and crop size of 128 instead of the defaults (64 and 96). Do you have any suggestions for how to tune these for different sets of data? The first changes I made seemed to work but I'd like to optimize for processing time if possible.

-Max

[YUNTAO LIU]

Hi Max,

Cube size of 32 and crop size of 128 in isonet predict is a good start point to try actually for all the dataset. 

Increasing crop_size and decreasing cube_size will greatly slow down the predicting step. The processing time can not be optimized if you need to do that. 

To view this discussion on the web visit https://groups.google.com/d/msgid/isonet/c554a73a-a1a3-4ff3-a288-df0f5589ed6cn%40googlegroups.com.

[Max Denena]

Got it. Thanks for the advice.