Generate mask images with Fiji or CellProfiler

Hi all, here is a new researcher in the community moving the first steps into the Image Analysis field !!

I assume a robust readout based on biological images heavily rely on segmentation procedure, so, as an initial step of my project, I would test a few algorithms on my images. Is out there any software that I can use to produce black&white masks to evaluate and compare algorithm? FIJI or CellProfiler maybe?!

Any feedback would be appreciated, thank you!

@darioxcarda
If you are just getting started - at least for ImageJ - here are some helpful links that will teach you what you need to know:

Start with the Workshops** - Introduction and Segmentation. That will definitely get you started. And if you have more specific questions regarding your own datasets - just let us know. We are here to help!

3 Likes

@darioxcarda

You can create binary mask in Cellprofiler too. If you are a beginner, you could try using the cellprofiler basic tutorials and example pipline to start with.
For creating binary images, you could use the module “Threshold” cellprofiler.

Regards,
Lakshmi
Fujifilm Wako Automation (Consultant)
www.wakoautomation.com
For Cellprofiler training or optimised pipeline write to,
lakshmi.balasubramanian.contractor@fujifilm.com

Join us at the High Content Analysis Symposium Oct 21-22, London
Society for Laboratory Automation and Screening

SLAS 2019 Advanced 3D Human Models and High-Content Analysis Symposium

Read more on our site.

Yokogawa CV8000 - The Ultimate in Confocal HCS
https://www.wakoautomation.com/products/yokogawa-high-content-imaging

2 Likes

Thank you for your inputs, I’m looking at Fiji to do initial thresholding.

I was looking for more interactive polygon-like segmentation, I found LabelMe (http://labelme.csail.mit.edu/Release3.0/) from MIT. Does anyone have already experience with this tool? Is out there a better solution? That ideally can be used on mobile devices.

I am looking for something to annotate a dataset of 50 images as ground truth for testing deep learning segmentation models, and eventually to compare expert-level annotations.

Any suggestion is welcome :slight_smile:

What type of ground truth will you be making? Full pixel outlines or bounding boxes or XY points?

If it’s the former, believe it or not I found Gimp to be the best solution! You can read a step by step here: https://blog.cellprofiler.org/2018/06/07/annotating-images-with-cellprofiler-and-gimp/

Then, to evaluate algorithms’ results you can use the MeasureObjectOverlap Module. Check out the help for that module in CellProfiler for more details!

2 Likes

Thank you so much that’s a very intuitive way of generating masks to test deep learning segmentation models. I just started with full pixel outlines, that in my case is very tedious, since the presence of tens of thousands of instances in each image.

Does anyone test an active learning strategy to refine an initial annotation of hundreds of instances? I was able to find only performance comparisons on the COCO dataset.

Thank you in advance!