Automating trainable WEKA segmentation on CLIJ2

Hi all!

When using trainable WEKA segmentation on Fiji, I generally automate the process by using the script available here. My images are 16-bit 3D stacks, each stack ranging from 6.5-10.5 MB in size, and my laptop has an NVIDIA GeForce GTX 1050 graphics card (3GB GDDR5). I am looking for some guidance on 1) Whether it is worthwhile to perform the analysis described thus on my GPU, and 2) If yes, how do I implement it with the help of CLIJ2?

Thanks in advance!

1 Like

Hey @mshah,

that’s a great question. Before we dive deeper, a comment: CLIJx-WEKA is experimental, that’s what the X stands for. So whatever you do with it, please be careful.

Your GPU memory is large enough and the GPU should be able to do it. To figure out if it’s worth, can you tell us know how long processing takes at the moment? Which features (+sigma range) did you select in Trainable Weka Segmentation?

For getting started, I’d recommend going through some basic CLIJ tutorials and the NEUBIAS Academy webinar about CLIJ2.

And before I start posting code snippets. Do you prefer programming Fiji in Jython or in Macro?


1 Like

Thanks for the response, @haesleinhuepf! I will check out the resources which you have linked.

I am experimenting with building models at the moment, so I use the entire sigma range with a minimum of 6-7 features (gaussian blur, mean, median, difference of gaussians, max, min, sobel, hessian and anisotropic diffusion). Depending on the final number of attributes and size of the stack, analysis takes anything between 1 minute - 10 minutes per 3D stack. I usually code in IJ Macro for all image analysis on Fiji.