Additional training data for neuronal subtypes from

https://pubmed.ncbi.nlm.nih.gov/36070775/ (Hamnet et al 2022)

specifically the Calretinin dataset

Check SPARC for updated datasets: https://sparc.science/data?type=dataset

A key feature that is requested is neuron segmentation in 3D.
It is possible, but as with any deep learning problem, generating a model requires manual annotation of data.
Challenges with 3D annotation:

• More effort annotating in 3D
• Variabiliity in image acquisition -> Anisotropic data ?.
• Image Quality: Variability in signal to noise ratio among images

The good news is that there is plenty of public data available now. Two main ones are:

• Mouse enteric neurons and subtypes: https://sparc.science/datasets/178/version/1
• Human enteric neurons and subtypes: https://sparc.science/datasets/55?type=dataset

This could be a community driven effort.. Perhaps make data available for annotation on a website. Apeer??

Once annotation is generated, it wouldn't be too challenging to generate a 3D segmentation model.

Currently the user selects the pixel size to rescale the images, so the cells to be segmented have the same size as the images used for training the model. This increases the accuracy of the segmentation.
Using the term pixel size can be confusing. Instead, we should use the term rescaling factor. Rescaling factor can be calculated by dividing the pixel size of the users image by 0.568 or whatever pixel size of the images the model has been trained on.

@mattyrowey

Could you ensure batch analysis option works and the documentation is clear?

You'll have to update GAT

Documentation is here:

https://gut-analysis-toolbox.gitbook.io/docs/2.-cell-counts-and-ganglia-analysis/batch-analysis

Is your feature request related to a problem? Please describe.
After analysis, all the csv results are scattered in different folders. Ryan Hamnett suggested that it would be a good idea to automatically collate them together.

Describe the solution you'd like
A script that we could point to an analysis directory that will collect all the csv files and make a final results file thats easier to handle

example macro on imagesc:

https://forum.image.sc/t/cvs-files-in-folder-into-one-excel-sheet-macro/37638
https://forum.image.sc/t/append-all-excel-files-xls-contained-in-a-folder/33288/2

Describe alternatives you've considered
Could look at using groovy instead of ImageJ macros for this if it gets complicated.

Add a cellpose model based on training data from GAT.
Default cyto2 model works quite well, but it can cause cell splitting.
Training with GAT data may produce a more accurate cellpose model.

Enable custom importing of ROIs for cells ; if using different software for prediction but want to use GAT workflow for measurements

Currently, the neuron subtype model is trained on both Hu and neuronal subtype annotations, and gives a lot of false positives.
The subtype model should only be trained on neuron subtype and not on Hu.

To do:

• Annotate new images to generate more training and test data
• Train a new model
• Validate the model on test data

• Add neuron subtype segmentation workflow in QuPath
• Add neuron segmentation , subtype segmentation using StarDist and downstream analysis using pyclesperanto assistant in napari

When I tick "image_already_open" in the test segmentation dialog, and I concurrently have a sample image open, the code terminates with a dialog popping up saying "'path' is required but unset." When unticked and I set the path to the image, this error does not happen. Perhaps the code is still pointing to a path, even when the image is opened already, and that checkbox is ticked? Thank you

!

The current pixel size used to train models is 0.568 micron per pixel.
When large images are used for analysis in GAT within FIJI, as the pixel size is quite high, they have to be resized to a higher resolution for appropriate segmentation. In essence, the image becomes larger and takes up more memory during the StarDist segmentation process. This is inefficient and should be changed in the future.
One solution is to increase the pixel size of images used for training the model..

Defining a ganglia is currently based on the presence of two markers, Hu and a neuronal/glial fibre marker.
This is not always practical and it takes up a valuable channel.
Proposal is to use Hu to define ganglia by expanding labels to a predefined distance. We can use the distances from spatial analysis as a starting points -> 6.5 um expansion of cell edges
Once expanded, use connected components labelling to define the ganglia.

• Define ganglia based on cell expansion and connected components
• Add a Test ganglia segmentation option under Tools to try different label expansion values

The ganglia segmentation model needs to be updated with new training data.
Also, to make training simpler and use better architecture, consider using this repo:
https://github.com/qubvel/segmentation_models

Get system specs and print it out to the Log Window

• CPU

• RAM

• GPU and specs

• Handle errors if no neurons detected

Describe the bug

Spatial Analysis returns neighbour count that is incremented by 1. This is because we are using the function "label overlap count map for calculating the number of neighbours after label dilation. It counts the overlapping cells and the same cell in the counts.
Currently, working on a bugfix to return the corrected values.

Will also add a feature to include spatial analysis as part of the standard workflow to minimize additional steps

PT-BIOP has a nice Fiji cellpose wrapper that calls cellpose from Fiji: https://github.com/BIOP/ijl-utilities-wrappers
Enable compatibility by allowing this as an option.
Points to consider:

• Instructions to install cellpose;

  • links to python environment config
  • video tutorial?

• Rescaling of image done in GAT for Stardist prediction. Determine the radius we pass to cellpose?

• Add test macro to test different options in Cellpose??

When using the GAT feature in Fiji, it allows me to count the number of neurons in the image I upload, but it will not count the number of neurons per ganglia. After it counts the neurons, and I okay what the program has labeled as a neuron, it gives me a Macro Error saying "Stack required in line 81 (called from line 64), Stack . ( ganglia_channel ) ;". When I click the "OK" button on this error. it displays a box that says "Ganglia image not a binary image". I am not really sure what any of this means, and my knowledge on this software is pretty limited. Is this something you are able to assist me on?