Counting PNNs with ImageJ

A short tutorial to help labmates to generate a multi-rater dataset for AI training

Why am I doing this?

Right… who didn’t ask this before during grad school?

Anyways, counting cells is tedious. There are ways to automate it, but PNNs are particularly tricky because they come in various shapes, sizes and intensities.

For this reason, we are building a neural network that can do the job for us. In order to evaulate the network training, we need a dataset of cells counted by an experimenter, but counting cells has some arbitrary parameters that depend on the experimenter.

If multiple experimenters count the same set of image we can build a dataset of PNNs which has, for each PNN, its probability to be counted.
This will help because we will be able to evaluate the importance of prediciton errors: making a wrong prediction on a PNN that have been counted by all the experimenters is terrible, while missing a PNN that have been catched by only 25% of the experimenters is not so bad.

What do I need?

For this task you will need to :

  • Download copy of ImageJ
  • Download the dataset of images to count (ask me privately for the drive link)

What do I need to do?

  1. Launch ImageJ (FIJI) and open one of the images that you have to count LoadFile

  2. Select the Multi-point tool. You will use this tool to count cells.

  3. Zoom to 100% by using the + key. It’s best to count cells at the same magnification factor, so let’s keep it 100%.

    • + : Zoom-in
    • - : Zoom-out
    • Spacebar + Drag : Pan around the image
    • Left-Click : Count a cell
    • Ctrl+Click : Delete a counted cell
  4. When you are done for this image, save the count.
    You have to open the ROI-manager, add the current selection to a ROI and then save the ROI.
    You should rename the file with the same name of the counted image but changing the extension from .tif to .roi

  5. You’re Done! When you have finished counting all the images, you will need to upload all the .roi files to a drive (contact me for the link).

Thanks!